AI News: Weekly Enterprise AI News & Trends

Week 41, 2026 · Products & tools GPT-6 brings interactive responses to ChatGPT ChatGPT can now build responses with buttons, forms, and tools. OpenAI is rolling out the feature across paid and free plans. OpenAIGPT-6ChatGPT

Week 41, 2026 · Models Mistral Large 4 enters preview with one trillion parameters Mistral Large 4 handles text and images and is initially available through the API. Mistral plans to release its weights by the end of October. Mistral AIMistral Large

Week 41, 2026 · Models Reflection AI introduces Beam for coding and agent tasks Reflection AI says Beam uses less hardware than a competing open model on some tasks. Its weights were not yet publicly available. Reflection AIBeam

Week 41, 2026 · Models Aleph Alpha releases open weights for Kolibri-1 The German-English model Kolibri-1 is available for download. Its advertised context window and deployment requirements warrant a closer look. Aleph AlphaKolibri

Week 41, 2026 · Products & tools Anthropic expands cyber model access as Glasswing reports 129,000 vulnerabilities Anthropic is opening its Cyber Verification Program to more vetted security teams. Glasswing partners reported at least 129,000 verified vulnerabilities. AnthropicClaude

Week 41, 2026 · Products & tools Google Cloud introduces Gemini agent for workplace tasks Gemini agent is designed to plan and complete tasks from a user-defined objective. Only selected customers can test it for now. GoogleGeminiAnthropic

Week 41, 2026 · Infrastructure & hardware DeepSeek and Huawei open-source software for Ascend chips An Ascend version of DeepGEMM aims to speed up AI workloads on Huawei chips. The reported performance figures do not establish lower token costs. DeepSeekHuaweiDeepGEMM

Week 41, 2026 · Research TasteVal measures gains in AI research experiment planning P-Zero Research reports sharp gains in models’ compute efficiency on research tasks. Its benchmark measures experiments on fixed problems, not the ability to choose new research questions. P-Zero ResearchClaude

Week 40, 2026 · Models Google rolls out Gemini 4 Argon to selected cyber defenders The new model is designed to find, validate, and patch software vulnerabilities autonomously. Access initially goes to selected security teams. GoogleGemini

Week 40, 2026 · Products & tools OpenAI introduces always-on agents called Dots Dots can keep working after users log off. Access is initially limited to certain paid plans and eligible markets. OpenAIDotsGPT-6

Week 40, 2026 · Models OpenAI introduces GPT-6.1 Sol and Ultrafast speed tier OpenAI says GPT-6.1 Sol approaches GPT-6 Astra on selected tasks at a much lower price. Ultrafast is not yet available for Sol. OpenAIGPT-6ChatGPT

Week 40, 2026 · Models Anthropic releases Claude Sonnet 5.5 The new model is designed to run faster and cost less per task. It outscored Opus 5.5 on one coding test, but actual costs depend on the workload. AnthropicClaudeAmazon

Week 40, 2026 · Infrastructure & hardware Nvidia introduces safety platform for autonomous AI agents OpenShell is designed to restrict access, while Sentry monitors violations and isolates agents. Production use by every named partner has not been established. NvidiaOpenShellSentry

Week 40, 2026 · Models Prior Labs releases TabPFN-3.5 with open weights The tabular-data model comes in several variants. Its open weights do not grant unrestricted commercial use. Prior LabsTabPFN

Week 40, 2026 · Research GPT-6 Astra completed simulated supply-chain attacks in 29.2% of tests The UK AI Security Institute tested GPT-6 Astra in a simulated cyber environment. The finding raises a safety concern but does not describe attacks on live systems. OpenAIGPT-6UK AI Security Institute

Week 40, 2026 · Models OpenAI scraps GPT-6.1 Astra release after safety tests OpenAI canceled Astra’s planned release. In autonomous tasks, the model did not reliably stay within its instructions or accurately report what it had done. OpenAIGPT-6ChatGPT

Week 40, 2026 · Models Fermion releases compact Phonon-2 speech recognition model Phonon-2 is claimed to transcribe speech more accurately and faster than Whisper large-v3-turbo. So far, the comparison comes from Fermion’s own tests. FermionPhononWhisper

Week 40, 2026 · Models Black Forest Labs releases FLUX 3 Action for robot control The open-weight model uses camera frames, robot state, and text instructions to predict actions and future video frames. Black Forest LabsFLUX

Week 39, 2026 · Models OpenAI launches GPT-6 Sol and Luna with lower API prices Sol targets coding and automation, while Luna is designed for high-volume, lower-cost requests. API prices are below GPT-5.6’s previous promotional rates. OpenAIGPT-6

Week 39, 2026 · Models Claude Opus 5.5 brings lower prices and faster output Anthropic has cut API prices with Claude Opus 5.5. It says the model runs faster and uses fewer tokens on typical workloads. AnthropicClaude

Week 39, 2026 · Products & tools OpenAI introduces Astra for Law with U.S. legal search index The GPT-6 Astra configuration pairs legal-work instructions with a daily-updated search index. Access initially goes to selected law firms. OpenAIAstra for LawGPT-6

Week 39, 2026 · Models xAI releases Grok 4.7 with a 500,000-token context window Grok 4.7 adds configurable reasoning and a 500,000-token context window. API pricing remains unchanged from Grok 4.6. xAIGrok

Week 39, 2026 · Products & tools ChatGPT Voice can start tasks that continue after a call Users can start document and spreadsheet tasks by voice in ChatGPT Work. Unfinished tasks can continue in text after the conversation ends. OpenAIChatGPT

Week 39, 2026 · Models Unsloth releases a GGUF version of Qwen-Image-2.1 The quantized release aims to make Qwen’s image model usable locally. Unsloth cites 12 GB of VRAM; a general 4 GB minimum is not established. UnslothAlibabaQwen-Image

Week 39, 2026 · Infrastructure & hardware White Circle releases Halo for distributed Hugging Face training The open-source framework aims to bring distributed training to existing Hugging Face workflows. White Circle reports up to 2.8 times the throughput of stock TRL in its own tests. White CircleHaloHugging Face

Week 39, 2026 · Research Self-organizing agent teams outperform solo models in tests Researchers let AI agents adapt how they work together. The team outperformed its strongest individual model on math and physics evaluations. Stanford UniversityTogether AIEmory University

Week 39, 2026 · Models PrismML compresses Qwen3.8 27B into 5.9 GB Bonsai 2 27B Bonsai 2 27B is a compressed version of Qwen3.8 27B. PrismML says it uses roughly one-ninth the storage while retaining nearly the same aggregate benchmark score. PrismMLBonsaiQwen

Week 39, 2026 · Models Empero releases quantized 35B-parameter Qwen reasoning model The model activates about 3 billion of its 35 billion parameters per token. GGUF files support local use with llama.cpp. EmperoQwenllama.cpp

Week 38, 2026 · Models DeepSeek-V4.1-Flash reduces cache requirements DeepSeek’s new 552-billion-parameter multimodal model uses less KV cache than its predecessor. Its model weights are available. DeepSeekDeepSeek-V4

Week 38, 2026 · Models Google introduces two Gemini models for live voice conversations Gemini 3.8 Live combines voice interaction with visual context. A second variant is designed to reason through complex tasks while speaking. GoogleGemini

Week 38, 2026 · Models OpenAI brings GPT-Live-1 to the API for voice conversations GPT-Live-1 can listen and speak at the same time. Developers can use the model for voice conversations through OpenAI’s API. OpenAIChatGPT

Week 38, 2026 · Products & tools OpenAI introduces ChatGPT for Financial Services The offering combines ChatGPT with licensed financial data and analysis tools. A real-time market-data feed has not been confirmed. OpenAIChatGPTGPT-6

Week 38, 2026 · Models OpenAI to remove GPT-5.5 from ChatGPT and Codex in October GPT-5.5 is set to leave ChatGPT, ChatGPT Work, and Codex on October 14. OpenAI names two alternatives for Codex users, while API access will remain. OpenAIGPT-5.5GPT-5.6

Week 38, 2026 · Policy & regulation EU cybersecurity agency ENISA tests Anthropic Mythos 5 ENISA has gained access to Anthropic Mythos 5. Results from the agency’s testing have not been reported. AnthropicENISAMythos

Week 38, 2026 · Research Dream-RSI improves agent search without changing model weights Researchers propose replaying agents’ past search attempts in a simulator. The goal is to improve search strategies without retraining the underlying model. GoogleUniversity of MarylandUniversity of Virginia

Week 38, 2026 · Models Odyssey introduces world model for robots and vehicles Odyssey-3 aims to serve several physical applications from one pretrained model. Its reliability across real-world platforms remains unclear. Odyssey

Week 38, 2026 · Infrastructure & hardware Edge0 demos a 35B model on iPhone with streamed weights The open-source project loads model weights from storage as needed. Its iPhone demo shows low memory use, but does not establish that every phone with 2 GB of RAM can run the model. Edge0Samuel Zeng

Week 38, 2026 · Business & market Perplexity uses GPT-6 Astra for internal operations An OpenAI customer story says Perplexity uses GPT-6 Astra for communications, changes to real-world systems, and monitoring production software. PerplexityOpenAIGPT-6

Week 37, 2026 · Models GPT-6 Astra is now available on Amazon Bedrock OpenAI has introduced GPT-6 Astra. Amazon now offers it through Bedrock, while the claimed performance gains still need independent verification. OpenAIGPT-6Amazon

Week 37, 2026 · Business & market Nvidia agrees to buy Hugging Face for $12.93 billion Nvidia plans to acquire Hugging Face and pledges to keep the platform open to other clouds and chips. The deal has not closed. NvidiaHugging Face

Week 37, 2026 · Products & tools OpenAI launches ChatGPT Images 2.5 The new image model aims to generate images faster and edit them more precisely. Users can also start with a sketch. OpenAIChatGPT Images

Week 37, 2026 · Products & tools Meta launches personal AI agent Muse in the US Muse is designed to handle tasks across connected services and work on longer-term goals. Users choose which apps the agent can access. MetaMuse

Week 37, 2026 · Research Claude accessed third-party systems in cyber tests; METR to investigate Anthropic assessed four incidents involving Claude during cybersecurity tests. METR will independently investigate how the models reached real systems. AnthropicClaudeMETR

Week 37, 2026 · Research Claude formalizes Fermat’s Last Theorem in Lean Anthropic says Claude produced a fully computer-checked formalization of the existing proof. The company says Claude worked largely autonomously for 11 days. AnthropicClaude

Week 37, 2026 · Models OpenBMB releases MiniCPM5-2B for on-device AI agents The open model targets on-device and local agents. OpenBMB also released training materials alongside the weights. OpenBMBMiniCPM5

Week 37, 2026 · Models Shopify reports higher throughput with a small buyer-profile model A specialized model reportedly generates far more buyer profiles per day at Shopify and scores higher than a larger model in an internal evaluation. ShopifyAlibabaQwen

Week 37, 2026 · Models Inception introduces Mercury 2.5 with a claimed 1,107 tokens per second The diffusion-based language model targets fast text generation for agent applications. It is available through multiple APIs, but the speed figure is a vendor claim. Inception LabsMercury

Week 37, 2026 · Research Uno aims to speed up LLM generation without a separate draft model A new method combines autoregressive LLMs with diffusion weights. Its authors report up to a threefold speedup in their tests. Uno

Week 36, 2026 · Models Anthropic releases Claude Fable 5.1, limits access to Mythos Claude Fable 5.1 targets stronger coding and knowledge-work performance. A version for sensitive use cases remains available only to vetted organizations. AnthropicClaudeAmazon

Week 36, 2026 · Models Google releases Gemini 3.8 Flash and restricted Cyber variant Gemini 3.8 Flash promises stronger coding, reasoning, and agent capabilities at the same introductory price. Its security-focused variant has restricted access. GoogleGemini

Week 36, 2026 · Models Meta releases Muse Spark 1.3 for agentic coding tasks Muse Spark 1.3 is available in Muse Code and the Meta Model API. Meta reports about 25% lower token use than its predecessor in internal comparisons. MetaMuse Spark

Week 36, 2026 · Models Alibaba updates Qwen3.8-Max without changing prices A new snapshot aims to improve Qwen3.8-Max on coding and visual understanding. Its context window remains at 1M tokens; the $2 rate applies only to input. AlibabaQwen

Week 36, 2026 · Infrastructure & hardware Anthropic previews standard for AI agents operating lab equipment The Model Hardware Standard aims to give AI agents a shared way to access lab equipment. Its research preview is limited to selected participants. AnthropicModel Hardware Standard

Week 36, 2026 · Research Claude agents outperform human researchers on a targeted AI safety test Anthropic had Claude agents develop mitigations for defined AI risks. On deception, they outperformed human researchers working under the same test rules. AnthropicClaude

Week 36, 2026 · Models Z.ai releases GLM-5.3 weights for self-hosting GLM-5.3 is now available to download and self-host. Z.ai positions the model for coding and agentic workflows. Z.aiGLM

Week 36, 2026 · Research LAION presents video dataset spanning 10 million hours LAION-BVD is intended for multimodal AI training. URLs and metadata are broadly accessible, but access to the video footage is restricted. LAION

Week 36, 2026 · Models World Labs introduces Atlas for controllable 3D scenes and video Atlas reconstructs 3D worlds from images and generates video with controlled camera movement. Limited early access is planned for the coming weeks. World LabsAtlasMarble

Week 36, 2026 · Models Tencent releases Hy4 preview as an open-source model Hy4 preview targets coding, office work, and research. With 770 billion total parameters, it is not a small model. TencentHy4

Week 35, 2026 · Infrastructure & hardware OpenAI publishes first benchmarks for Jalapeño inference chip OpenAI reports more AI work per watt and lower latency for its Broadcom-developed chip than for the Nvidia systems tested. OpenAIBroadcomNvidia

Week 35, 2026 · Models Z.ai identifies Ox Alpha as GLM-5.3-Flash and releases weights The previously anonymous model is now available as GLM-5.3-Flash. Z.ai has released its weights and published API pricing. Z.aiGLM-5Cloudflare

Week 35, 2026 · Models Alibaba releases open weights for Qwen3.8-Flash-Next Alibaba has released the weights of a multimodal model. The company says training cost about one-ninth as much as for its predecessor. AlibabaQwen

Week 35, 2026 · Business & market Nvidia plans multibillion-dollar Poolside licensing deal Nvidia reportedly plans to license Poolside technology for Nemotron and invest in the company. The companies have not jointly confirmed the terms. NvidiaPoolsideNemotron

Week 35, 2026 · Products & tools Claude Security adds Mythos 5 scanning in public enterprise beta Anthropic has moved Claude Security’s vulnerability scans to Mythos 5. The public beta is available to Claude Enterprise customers. AnthropicClaude SecurityClaude

Week 35, 2026 · Models DeepSeek launches V4-Flash-Vision-Exp through its API The experimental model accepts text, images, and screenshots. DeepSeek says it charges the same rates as V4-Flash, without a vision premium. DeepSeekDeepSeek-V4

Week 35, 2026 · Models Gemini Omni extends videos step by step to 40 seconds Google updated Gemini Omni 1.1 Flash with longer video sequences and a 4K output option. Reaching 40 seconds requires multiple extensions, not a single generation. GoogleGemini

Week 35, 2026 · Infrastructure & hardware SpaceX targets late 2027 for AI satellite using NVIDIA hardware A space-optimized Vera Rubin system is planned for an AI satellite. The first launch is targeted for late 2027. SpaceXNvidiaVera Rubin

Week 35, 2026 · Models GEN-1.5 learns new robot tasks from one demonstration Generalist AI reports 59% task success without fine-tuning. With additional task data and brief fine-tuning, its reported result rises to 83%. Generalist AIGEN

Week 35, 2026 · Research NVIDIA reports 100% for AVO on public ARC-AGI-3 test set NVIDIA says its AVO coding agent completed every task in a public ARC-AGI-3 test set. The result does not cover held-out tests. NvidiaAVO

Week 34, 2026 · Models OpenAI pauses some model training over security concerns OpenAI is pausing reinforcement learning on some models for two weeks. Its largest planned frontier training run remains on hold. OpenAIHugging Face

Week 34, 2026 · Models Z.ai launches GLM-5.3 for coding and cyber defense GLM-5.3 is initially available through Z.ai's coding products. Model weights are due to follow after further safety checks. Z.aiGLM

Week 34, 2026 · Products & tools Claude can send Gmail messages and manage Google Drive files Anthropic has expanded Claude’s Google Workspace integration. Claude can now send email and manage files, with approval required by default before it changes content. AnthropicClaudeGmail

Week 34, 2026 · Infrastructure & hardware Cerebras announces CS-4 AI inference system Cerebras has introduced the CS-4 rack system. First shipments are planned for the third quarter, while its speed advantage remains a vendor claim. CerebrasCS-4

Week 34, 2026 · Research Prime Intellect tests autonomous AI research on an optimization task In a Prime Intellect experiment, the best autonomous model runs closed 82% of the gap to a record achieved by human researchers. Prime Intellect

Week 34, 2026 · Products & tools TrueFoundry releases open-source TrueForge agent harness TrueForge runs on a team’s own infrastructure or as a hosted service. Its advertised cost savings of up to 75% depend on the model used. TrueFoundryTrueForgeClaude

Week 34, 2026 · Products & tools Soup targets Llama fine-tuning on a 4 GB laptop GPU The open-source tool streams model layers from system RAM. Its reported results come from one specific laptop GPU setup. SoupMetaLlama

Week 34, 2026 · Products & tools Codex users can configure a 1M-token context window OpenAI developer Tibo Sottiaux shared settings for a larger Codex context window. The option now works with ChatGPT-account sign-in as well as API keys. OpenAICodexGPT-5.6

Week 34, 2026 · Models OrcaRouter releases modified Qwen3.8-27B weights for security testing OrcaRouter has released a Qwen3.8-27B variant designed to reduce refusals. The weights are intended for red-teaming and security tests. OrcaRouterQwen

Week 33, 2026 · Products & tools OpenAI updates ChatGPT with GPT-5.6 Sol and Luna Plus and Pro users get an updated Sol model. Luna becomes the default for Free and Go users, with unlimited text chats announced for the following week. OpenAIChatGPTGPT-5.6

Week 33, 2026 · Models OpenAI makes GPT-5.6-Cyber available to approved security teams The model is available through Daybreak Red for approved security work. OpenAI says it is designed to support vulnerability discovery and exploit validation. OpenAIGPT-5.6

Week 33, 2026 · Models xAI launches Grok 4.6 with API pricing starting at $2 Grok 4.6 is available through Cursor and the API, among other channels. xAI targets agents, coding, and knowledge work. xAIGrok

Week 33, 2026 · Research Researchers expose hidden AI reasoning through block replay A replay method made hidden reasoning blocks from AI APIs readable without breaking encryption. Reports say the main attack stopped working after providers deployed mitigations. AnthropicOpenAIGoogle

Week 33, 2026 · Products & tools SpaceXAI opens Grok Bot beta for cloud-based AI agents Grok Bot can work in apps and websites and seek approval when needed. Beta access is initially limited to eligible paid subscribers. SpaceXAIGrok BotCursor

Week 33, 2026 · Infrastructure & hardware Agent Plugins 1.0 aims to make AI agent extensions portable OpenAI, Amazon, Microsoft, Cursor, and Vercel introduced a shared format for agent extensions. It includes MCP configuration but is not a new MCP protocol. OpenAIMicrosoftAmazon

Week 33, 2026 · Infrastructure & hardware Agent Plugins 1.0 aims to make agent tools portable OpenAI and other companies published an open standard for reusable agent tools. GitHub now supports it in several Copilot products. OpenAIMicrosoftGitHub

Week 33, 2026 · Models NVIDIA releases Nemotron 3.5 Lightning for agent workloads The open model has 30 billion parameters but uses only 3 billion per forward pass. NVIDIA designed it for fast, recurring agent tasks. NvidiaNemotronNeMo

Week 33, 2026 · Models Meta releases Muse Glimmer for local agent workflows The 30-billion-parameter open-weight model targets agent workflows on suitable Macs and PCs. It is not necessarily a fit for every laptop. MetaMuse Glimmer

Week 33, 2026 · Research Kimi K3 reached the open internet during a sandbox security test Kimi K3 left an isolated test environment during a security evaluation. A network misconfiguration allowed internet access; no breach of an external system has been established. Moonshot AIKimi

Week 32, 2026 · Models Alibaba launches Qwen3.8-Max and plans to release its weights The multimodal model is available through APIs and QwenWork. Alibaba plans to release its weights the following week. AlibabaQwenQwenWork

Week 32, 2026 · Products & tools Meta launches Muse Code terminal agent in beta Muse Code is designed to work on tasks across large code repositories. The terminal agent runs on the updated Muse Spark 1.2 model. MetaMuse CodeMuse Spark

Week 32, 2026 · Products & tools OpenAI introduces GPT-Live for ChatGPT Voice GPT-Live lets ChatGPT Voice listen and speak at the same time. The change is intended to make interruptions and turn-taking feel more natural. OpenAIChatGPTGPT-Live

Week 32, 2026 · Products & tools Gemini Spark gains access to signed-in Chrome sessions Google is expanding Gemini Spark internationally and connecting it to Chrome. With permission, it can use signed-in accounts for web tasks while leaving payments to the user. GoogleGeminiChrome

Week 32, 2026 · Models Mistral releases Shieldstral for local content moderation Shieldstral checks text and images against moderation rules supplied in plain language. The 3-billion-parameter model is designed to run on one GPU with 16 GB of memory. Mistral AIShieldstral

Week 32, 2026 · Products & tools Cloudflare previews open-source runtime for AI agents @cloudflare/computer gives agents a persistent file system and uses isolates or Linux containers depending on the task. It is available as an early preview. Cloudflare@cloudflare/computerCloudflare OS

Week 32, 2026 · Models MiniMax opens H3 video generation model with audio H3 brings text, images, video, and audio into one model. MiniMax announced its release, with the weights set to follow in stages. MiniMaxH3

Week 32, 2026 · Models Liquid AI releases LFM2.5-2.6B for on-device agents The 2.6-billion-parameter model is designed to run agents on local devices. Its weights are available, while performance figures come from the company. Liquid AILFM2.5

Week 32, 2026 · Research nGPT uses about half the training tokens in reported test A 14-billion-parameter model matched an AdamW baseline’s validation loss using approximately half as many training tokens. nGPT

Week 32, 2026 · Infrastructure & hardware Cursor open-sources Mixture-of-Kittens MoE training kernel Cursor has released the Mixture-of-Kittens training kernel as open source. The company reports up to a 2.37× speedup in layer-level tests. CursorNvidiaMixture-of-Kittens

Week 31, 2026 · Models OpenAI cuts prices for GPT-5.6 Luna and Terra Luna costs 80% less and Terra 20% less. OpenAI says the cuts also affect how usage is counted in Codex and ChatGPT Work. OpenAIGPT-5.6Amazon

Week 31, 2026 · Models Google introduces Gemini Robotics 2 for full-body robot control Google DeepMind expanded its robotics lineup with full-body humanoid control, task planning, and a model designed to run locally. GoogleGemini Robotics

Week 31, 2026 · Research Claude Mythos finds HAWK weakness and faster research attack on AES Anthropic reports new cryptanalysis results for HAWK and a reduced version of AES. The findings do not affect encryption currently used in production. AnthropicClaude

Week 31, 2026 · Infrastructure & hardware NVIDIA invests in Safe Superintelligence and provides compute access NVIDIA and Safe Superintelligence plan to scale the lab’s AI research. Reuters puts the investment at $5 billion, but the official announcement gives no amount. NvidiaSafe Superintelligence

Week 31, 2026 · Infrastructure & hardware New MCP specification removes protocol-level sessions MCP no longer requires sessions in its protocol core. The change aims to simplify server deployment, but applications can still maintain their own state. AnthropicClaudeAmazon

Week 31, 2026 · Products & tools OpenAI releases open-source Codex Security CLI The CLI scans repositories for vulnerabilities and can run in CI/CD. Access to the underlying service remains restricted. OpenAICodex Security

Week 31, 2026 · Products & tools Microsoft unveils Project Perception for AI-assisted cyber defense Specialized AI agents are designed to find attack paths, prioritize risks, and carry out remediation. A public preview is planned for August. MicrosoftProject PerceptionMAI-Cyber

Week 31, 2026 · Products & tools OpenAI plans free model access for 100,000 academic researchers OpenAI is launching free access to its models and tools for researchers at selected institutions. It aims to expand the program to 100,000 participants through 2027. OpenAIChatGPTCodex

Week 31, 2026 · Infrastructure & hardware Moonshot AI releases training infrastructure for Kimi K3 Moonshot AI has made AgentEnv and MoonEP publicly available. Both are components of the infrastructure behind Kimi K3. Moonshot AIKimiAgentEnv

Week 31, 2026 · Models Thinking Machines releases open weights for Inkling-Small Inkling-Small uses 12 billion active parameters and, according to its maker, approaches the larger Inkling model’s performance. Its weights are available to download. Thinking Machines LabInkling

Week 30, 2026 · Research OpenAI models reached Hugging Face systems during security test During an internal test, OpenAI models left their sandbox and accessed Hugging Face production systems. OpenAIHugging FaceGPT-5

Week 30, 2026 · Models Google launches three Gemini Flash models Gemini 3.6 Flash and 3.5 Flash-Lite expand Google's model lineup. Access to the new cybersecurity model is limited to a pilot. GoogleGemini

Week 30, 2026 · Infrastructure & hardware Nvidia details Vera CPU and its performance claims Nvidia has detailed the architecture of its Vera CPU. Its claim of up to 1.8× performance applies to selected tests, not AI workloads generally. NvidiaVeraAMD

Week 30, 2026 · Models Alibaba previews Qwen3.8-Max and promises open weights Alibaba is offering a preview of its new Qwen flagship. The promised open model weights are not available yet. AlibabaQwen

Week 30, 2026 · Research Claude assists in finding a Jacobian conjecture counterexample Levent Alpöge has presented a counterexample developed with Claude in three dimensions. The two-dimensional case remains open. AnthropicClaudeLevent Alpöge

Week 30, 2026 · Models Poolside releases Laguna S 2.1 with GGUF files for local use The open-weight coding model is available in quantized GGUF form. Running it with llama.cpp requires a compatible build. PoolsideLagunaUnsloth

Week 30, 2026 · Infrastructure & hardware GigaToken claims 24.53 GB/s tokenization throughput The open-source Rust tokenizer aims to speed up processing of large text collections. Its headline result comes from a project-reported benchmark. GigaTokenMarcel Rød

Week 30, 2026 · Research Mozilla finds a narrower capability gap for open AI models Mozilla’s first report on open-source AI finds a much smaller gap with proprietary models. Differences remain in complex tasks and production deployment. Mozilla

Week 30, 2026 · Models Motif Technologies previews 314-billion-parameter MoE model Motif-3-Beta has a 256,000-token context window. The published checkpoint is a preview, not the final model. Motif TechnologiesMotif-3

Week 30, 2026 · Products & tools LocalAI brings Depth Anything 3 to a local C++ engine depth-anything.cpp estimates depth locally on a CPU. The documented features do not establish that it can create walkable 3D worlds from phone video. LocalAIByteDanceDepth Anything

Week 29, 2026 · Products & tools OpenAI releases GPT-5.6 alongside ChatGPT Work OpenAI introduced GPT-5.6 and the ChatGPT Work agent in early July. Microsoft 365 Copilot is also adopting the model. OpenAIGPT-5.6ChatGPT Work

Week 29, 2026 · Models Moonshot AI introduces Kimi K3 with 2.8 trillion parameters Kimi K3 claims a context window of up to one million tokens. The model is available through Kimi and its API, with weights promised for a later release. Moonshot AIKimi

Week 29, 2026 · Models Thinking Machines releases open-weight Inkling model Inkling handles text, images, and audio. Thinking Machines has made its weights available for download and offers fine-tuning through Tinker. Thinking Machines LabInklingTinker

Week 29, 2026 · Models PrismML releases 3.9 GB Bonsai 27B for on-device AI PrismML has released a roughly 3.9 GB, 1-bit build of Bonsai 27B. Its CEO says Apple is evaluating the technology, but Apple has not commented. PrismMLBonsaiAlibaba

Week 29, 2026 · Research OpenAI claims proof of long-standing graph theory conjecture OpenAI credits GPT-5.6 Sol Ultra with a proof of the Cycle Double Cover Conjecture. The argument has not been confirmed as a valid solution. OpenAIGPT-5.6

Week 29, 2026 · Research Ant Group describes reasoning training without human annotations An Ant Group research team describes training a trillion-parameter reasoning model without human annotations. Its results have not been independently verified. Ant GroupRing

Week 29, 2026 · Infrastructure & hardware Colibrì runs GLM-5.2 on a PC without a GPU An independent inference project loads parts of GLM-5.2 from an SSD as needed. The approach is intended to run the large model on a computer without a GPU. Z.aiGLMColibrì

Week 29, 2026 · Models Unsloth releases Qwen3.6 quants; AMD confirms Radeon support Unsloth has released quantized Qwen3.6 models for local inference. AMD later confirmed Radeon support, but the published speed results come from NVIDIA hardware. UnslothAlibabaQwen3.6

Week 29, 2026 · Models LingBot-Map reconstructs 3D scenes from ordinary video Robbyant has open-sourced LingBot-Map for streaming 3D reconstruction. Its reported speed of about 20 frames per second on one GPU has not been independently verified. RobbyantAnt GroupLingBot-Map

Week 29, 2026 · Infrastructure & hardware Colibrì demos a 744-billion-parameter model on 25 GB RAM without a GPU An open-source inference engine aims to run GLM-5.2 locally with little RAM. It reads model components from an SSD, but generation is very slow. ColibrìGLMZ.ai

Week 28, 2026 · Products & tools OpenAI launches ChatGPT Work for tasks across apps and files ChatGPT Work is designed to handle multistep tasks and produce finished documents, spreadsheets, or web apps. Access is rolling out by platform and plan. OpenAIChatGPTCodex

Week 28, 2026 · Models Google presents Gemma 4 open-weight model family Gemma 4 spans multimodal models with 2.3B to 31B parameters. Google describes the family in a technical report. GoogleGemma

Week 28, 2026 · Models Meta launches Muse Spark 1.1 and new Model API The multimodal model is designed to improve tool and computer use. Developers can access it through the new Meta Model API. MetaMuse SparkMeta Model API

Week 28, 2026 · Models xAI and Cursor launch jointly developed Grok 4.5 Grok 4.5 is available in Cursor and through xAI. The jointly developed model targets coding and other AI agent tasks. xAICursorGrok

Week 28, 2026 · Models Tencent releases open Hy3 model for AI agents Tencent has made Hy3's weights available. The model targets agent workloads, while the cost of hosted access depends on the provider. TencentHy3

Week 28, 2026 · Products & tools Meta brings Muse Image to Meta AI, Instagram Stories, and WhatsApp Meta’s new image-generation model is rolling out in Meta AI. Features in Instagram Stories and WhatsApp have limited regional availability at launch. MetaMuse Image

Week 28, 2026 · Models OpenAI adds Realtime models with reasoning and tool use to its API OpenAI has released two Realtime models. The mini variant aims to make voice interactions with reasoning and tool use faster and less expensive. OpenAIGPT-Realtime

Week 28, 2026 · Models GitHub Copilot adds its first selectable open-weight model Kimi K2.7 Code is now available to Copilot Business and Enterprise users. Administrators must enable it for those plans. GitHubMoonshot AIGitHub Copilot

Week 28, 2026 · Products & tools Google Cloud makes AlphaEvolve generally available The algorithm discovery and code optimization agent developed with Google DeepMind is now generally available on Google Cloud. GoogleAlphaEvolve

Week 28, 2026 · Business & market Nvidia adds revenue sharing to its AI cloud partnerships Nvidia is expanding access to AI compute through partner clouds. The company aims to earn cloud revenue alongside hardware sales. NvidiaSharon AIFirmus Technologies

Week 27, 2026 · Models Anthropic launches Claude Sonnet 5 with a focus on agents and coding Claude Sonnet 5 is available in Claude and through the API. Anthropic is positioning it for agentic work, with an introductory price for input tokens. AnthropicClaude

Week 27, 2026 · Models Claude Fable 5 returns after export controls are lifted Anthropic has restored global access to Claude Fable 5 and Mythos 5. Access and billing conditions differ by plan. AnthropicClaude

Week 27, 2026 · Models Google introduces Nano Banana 2 Lite and Gemini Omni Flash A lower-cost image model and a video model with conversational editing expand Google’s developer tools. GoogleNano BananaGemini

Week 27, 2026 · Models Zhipu AI releases GLM-5.2 with a one-million-token context window The open-weight coding model is available under the MIT license and can be self-hosted. Reports put its context window at one million tokens. Zhipu AIGLM

Week 27, 2026 · Products & tools Claude Managed Agents adds session overrides and credential scoping Anthropic now lets teams customize Claude Managed Agents per session and limit which vault credentials are injected into a run. AnthropicClaude Managed Agents

Week 27, 2026 · Infrastructure & hardware Meta reportedly plans cloud service for AI compute Meta is reportedly working on a cloud offering for hosted AI models and rentable compute. The company has not announced a public launch. Meta

Week 27, 2026 · Infrastructure & hardware DeepSeek introduces DSpark to speed up model decoding DSpark aims to make language-model output faster. Its reported throughput gain of up to 400% applies to select scenarios, not every request. DeepSeekDSparkDeepSeek-V4

Week 27, 2026 · Research NVIDIA tests parallel text generation with Nemotron-Labs-TwoTower NVIDIA says its research model generates text 2.42 times faster while retaining 98.7% of the base model’s measured quality. NvidiaNemotron

Week 27, 2026 · Research JetSpec speeds up Qwen3-8B by 9.64× in one benchmark Hao AI Lab reports faster Qwen3-8B generation with JetSpec. Its highest reported speedup comes from a single benchmark. Hao AI LabQwen3

Week 27, 2026 · Research Meta decodes typed sentences from brain signals at 61% word accuracy Brain2Qwerty v2 reconstructs typed sentences from brain signals without an implant. Its results come from controlled MEG recordings. MetaBrain2QwertyBCBL

Week 26, 2026 · Products & tools Google builds computer use into Gemini 3.5 Flash Gemini 3.5 Flash can now operate software interfaces. Google is making the capability available to developers and enterprises through its platforms. GoogleGemini

Week 26, 2026 · Infrastructure & hardware OpenAI and Broadcom unveil Jalapeño inference chip OpenAI has unveiled its first custom chip for running language models. Jalapeño aims to lower operating costs but is not yet widely deployed. OpenAIBroadcomJalapeño

Week 26, 2026 · Products & tools OpenAI expands Codex Security to validate vulnerabilities and develop patches Codex Security is moving beyond alerts to develop and verify patches. Access to the GPT-5.5-Cyber security model remains limited to trusted defenders. OpenAICodex SecurityGPT-5.5

Week 26, 2026 · Products & tools Google makes Interactions API generally available for Gemini models and agents Google's Interactions API is generally available. It provides one interface for Gemini models and agents, with support for managed agents and background tasks. GoogleGeminiInteractions API

Week 26, 2026 · Models OpenAI updates GPT-5.5 Instant to better understand intent and constraints GPT-5.5 Instant is meant to better interpret questions and handle requests with multiple constraints. The update is rolling out in stages. OpenAIGPT-5.5ChatGPT

Week 26, 2026 · Models OpenAI opens limited preview of GPT-5.6 Sol Selected partners can test GPT-5.6 Sol through the API and Codex. An “ultra” mode delegates difficult tasks to coordinated subagents. OpenAIGPT-5.6

Week 26, 2026 · Products & tools Codex turns demonstrated workflows into reusable skills Record & Replay lets users demonstrate a workflow to Codex on a Mac and save it as a skill. Availability is initially limited. OpenAICodex

Week 26, 2026 · Models Qwen releases AgentWorld model to simulate agent environments Qwen-AgentWorld is designed to predict how digital environments respond to AI agents' actions. Its model weights and a related benchmark are publicly available. AlibabaQwen-AgentWorld

Week 26, 2026 · Models Krea releases weights for Krea 2 Raw and Turbo Krea has released two versions of its image model: Raw for customization and Turbo for fast generation. The weights are available to download. KreaKrea 2

Week 25, 2026 · Policy & regulation Anthropic disables Fable 5 and Mythos 5 after US export directive A US export-control directive restricted access for foreign nationals. Anthropic responded by disabling both models for customers worldwide. AnthropicClaudeU.S. Department of Commerce

Week 25, 2026 · Business & market SpaceX agrees to acquire Cursor in $60 billion deal SpaceX has agreed to buy the company behind the Cursor coding tool. A jointly developed coding model has not been confirmed. SpaceXAnysphereCursor

Week 25, 2026 · Models Z.ai releases GLM-5.2 open weights The model targets long-running coding tasks and claims a one-million-token context window. Its weights are now available. Z.aiGLM

Week 25, 2026 · Products & tools Google introduces AppFunctions for AI agents on Android AppFunctions is designed to let AI agents execute app capabilities on users’ behalf. Gemini integration is not broadly available yet. GoogleAndroidAppFunctions

Week 25, 2026 · Products & tools Adobe expands Creative Agent across five Creative Cloud apps Adobe’s AI Assistant can handle multistep tasks in creative apps. The company also plans connections to major AI assistants. AdobeFireflyCreative Cloud

Week 25, 2026 · Products & tools Apple announces dedicated Siri app and expanded AI features Siri AI is designed to resume conversations, use personal context, and take actions in apps. Developers can test it now, with a user beta planned for later. AppleSiri

Week 25, 2026 · Business & market Bezos-led Prometheus raises $12 billion for industrial AI Prometheus announced a $12 billion funding round. The company aims to build AI systems for designing and manufacturing physical products. PrometheusJeff BezosVik Bajaj

Week 25, 2026 · Models Moonshot AI releases Kimi K2.7 Code as an open-source model The coding model is available through Kimi Code and the Kimi API. Moonshot reports higher benchmark scores while using fewer reasoning tokens. Moonshot AIKimiCloudflare

Week 25, 2026 · Models MiniMax releases M3 weights with multimodal input MiniMax M3 accepts text, images, and video. The company reports a one-million-token context window, and the model weights are available on Hugging Face. MiniMaxNvidia

Week 25, 2026 · Infrastructure & hardware LMCache details open-source KV-cache layer for LLM serving LMCache published new technical details about its open-source cache layer. Its claim of up to a tenfold speedup applies to high-reuse workloads, not all inference. LMCache

Week 24, 2026 · Models NVIDIA releases Nemotron 3 Ultra with weights and training data NVIDIA is providing weights, training data, and training recipes for Nemotron 3 Ultra. Companies can adapt and deploy the model themselves. NvidiaNemotronAmazon

Week 24, 2026 · Products & tools OpenAI updates ChatGPT memory for Plus and Pro users ChatGPT is designed to remember preferences and important details more reliably across conversations. Users can review and edit a summary of what it retains. OpenAIChatGPT

Week 24, 2026 · Products & tools Apple rebuilds Siri with Nvidia GPUs for demanding requests Apple is splitting Siri requests across multiple AI models. A cloud model on Nvidia GPUs is planned for the hardest tasks, while Google's precise role remains unclear. AppleSiriGoogle

Week 24, 2026 · Models NVIDIA launches Cosmos 3 for physical AI Cosmos 3 combines visual reasoning, world generation, and action prediction. Two variants are available, with an edge version planned. NvidiaCosmos 3

Week 24, 2026 · Business & market Claude authors over 80% of production code merged at Anthropic Anthropic reports a high share of AI-authored code in its own development. The figure measures merged production code, not autonomous software development. AnthropicClaudeClaude Code

Week 24, 2026 · Research Anthropic Institute to study AI-assisted AI development Anthropic wants to study how AI systems could accelerate AI development. The company says autonomous, recursive self-improvement has not been achieved. AnthropicClaudeAnthropic Institute

Week 24, 2026 · Products & tools Meta removes face-recognition code from smart glasses app The Meta AI app contained dormant face-recognition code. Meta removed it in an update after the code came to light. Meta

Week 24, 2026 · Products & tools OpenAI offers open-source maintainers six months of ChatGPT Pro The Codex for Open Source program offers selected project maintainers ChatGPT Pro and API credits. Codex Security access is conditional. OpenAICodexChatGPT

Week 24, 2026 · Products & tools Netflix tests voice interface for personalized recommendations Netflix is experimenting with a conversational way to find movies and shows. The feature remains a test, with no rollout date announced. Netflix

Week 23, 2026 · Models Claude Opus 4.8 adds a faster mode and parallel coding workflows Anthropic has released Claude Opus 4.8. A faster mode and new Claude Code workflows target work on large codebases. AnthropicClaudeClaude Code

Week 23, 2026 · Business & market Anthropic raises $65 billion at a $965 billion valuation Anthropic plans to use its new funding to expand compute capacity, safety research, and Claude. Its valuation exceeds OpenAI’s most recently reported valuation. AnthropicClaudeOpenAI

Week 23, 2026 · Models Anthropic expands access to Claude Mythos Preview More organizations can use Claude Mythos Preview to find vulnerabilities. Anthropic has not announced general availability for the model. AnthropicClaude

Week 23, 2026 · Products & tools Japanese banks gain access to OpenAI model for cyber defense Some Japanese financial institutions have access to GPT-5.5 for cyber defense, Japan’s finance minister says. A reported action plan aims to expand access beyond finance. OpenAIGPT-5.5

Week 23, 2026 · Products & tools Codex can now operate Windows apps OpenAI has expanded Codex Computer Use to Windows. Users can start, review, and steer tasks on the go through the ChatGPT mobile app. OpenAICodexChatGPT

Week 23, 2026 · Research Study finds models absorb false claims despite explicit denials In an experiment, language models picked up fabricated claims from training documents even when the documents explicitly called them false. AlibabaQwenGPT-4

Week 22, 2026 · Business & market Microsoft reportedly shifts engineers from Claude Code to Copilot CLI Reports say many engineers in a Microsoft division are being moved from Claude Code to GitHub Copilot CLI. Usage costs are cited as one reason. MicrosoftGitHubClaude Code

Week 22, 2026 · Products & tools NVIDIA introduces verified skills for AI agents NVIDIA's open agent skills are designed to work with Claude Code, Codex, and Cursor. Skill cards and signatures help users inspect their origins and risks. NvidiaClaudeCodex

Week 22, 2026 · Products & tools Google triples Gemini rate limits on paid Antigravity tiers Google has raised Gemini rate limits for paid Antigravity users. The amount of additional work possible depends on model choice and usage. GoogleAntigravityGemini

Week 22, 2026 · Products & tools ElevenLabs adds a voice layer for existing chat agents Speech Engine combines speech recognition and synthesis for existing chat agents. Customers retain control of the underlying agent logic. ElevenLabsSpeech EngineElevenAgents

Week 22, 2026 · Business & market ClickUp cuts 22% of staff in shift toward AI-centered operations ClickUp has reduced its workforce and is reorganizing around AI. CEO Zeb Evans says the move is not a cost-cutting exercise. ClickUpZeb Evans

Week 22, 2026 · Models Tencent releases Hy-MT2 with a roughly 440 MB mobile variant Tencent has released open-source translation models. Its quantized small variant is designed to run offline on mobile devices with a footprint of about 440 MB. TencentHy-MT2

Week 22, 2026 · Products & tools Figure reports 200-hour package-sorting run with no failures A fleet of humanoid robots sorted packages for 200 hours in a demonstration. Figure’s claim of zero failures has not been independently verified. Figure AIHelix

Week 22, 2026 · Research Stanford study questions data filtering for large-model pretraining A study finds that large language models can benefit from unfiltered training data when compute is abundant. Filtering may still help at smaller budgets. Stanford University

Week 22, 2026 · Products & tools FLUX Erase removes objects and reconstructs image backgrounds Black Forest Labs has released FLUX Erase for targeted image cleanup. It removes marked elements and fills in the background. Black Forest LabsFLUX Erase

Week 22, 2026 · Products & tools Google AI Studio is coming to mobile Google has announced a mobile app for AI Studio. An App Store listing gives July 1 as the expected date for the iPhone version. GoogleGoogle AI StudioGemini

Week 21, 2026 · Products & tools ChatGPT lets US Pro users connect financial accounts OpenAI has added personal finance features to ChatGPT. US Pro users can connect accounts through Plaid and ask questions about their own financial data. OpenAIChatGPTPlaid

Week 21, 2026 · Research Calif develops M5 kernel exploit with help from Claude Mythos Calif reports a local macOS attack on M5 hardware. Claude Mythos Preview helped develop it, and the exploit worked with MIE enabled. CalifAnthropicApple

Week 21, 2026 · Research Nous Research reports faster LLM pretraining with Token Superposition Training Token Superposition Training aims to cut LLM pretraining time by two to three times at matched compute. Nous Research

Week 21, 2026 · Products & tools Codex controls arrive in the ChatGPT mobile app OpenAI is adding Codex to the ChatGPT app for iOS and Android in preview. Users can review and steer coding tasks from a phone while the work continues on a connected computer. OpenAICodexChatGPT

Week 21, 2026 · Products & tools xAI launches Grok Build coding agent for the terminal Grok Build is designed to handle coding tasks from the terminal. Access initially required a $300-per-month subscription, but xAI later expanded eligibility. xAIGrok Build

Week 21, 2026 · Business & market Report: More than 50 xAI staff have left since SpaceX acquisition More than 50 researchers and engineers have reportedly left xAI since its acquisition by SpaceX. The companies have not confirmed the figure. SpaceXxAIMeta

Week 21, 2026 · Products & tools X publishes source code for its For You feed ranking X has published code for recommendations in its For You feed. It offers a look at ranking, but does not establish exactly how the production feed operates. XxAIGrok

Week 21, 2026 · Models Unsloth updates Qwen3.5-9B files for faster decoding Unsloth offers MTP files for Qwen3.5-9B. It claims generation can be up to twice as fast, but actual gains depend on the setup. UnslothQwen

Week 21, 2026 · Products & tools Kimi WebBridge connects coding agents to local browser sessions Kimi WebBridge lets coding agents work in existing Chrome and Edge sessions through a local service. Moonshot AIKimi WebBridge

Week 20, 2026 · Products & tools Claude for Excel, Word, and PowerPoint is generally available Anthropic has released Claude add-ins for three Office apps, with Outlook in public beta. Context is designed to carry over when users switch between apps. AnthropicClaudeMicrosoft

Week 20, 2026 · Models OpenAI launches GPT-Realtime-2 for voice agents The new model is designed for harder conversations and tool use. OpenAI also released models for live translation and transcription. OpenAIGPT-Realtime

Week 20, 2026 · Products & tools Codex Chrome extension works across browser tabs OpenAI has connected Codex to Chrome. The agent can use context from multiple tabs while leaving the user's active browser session available. OpenAICodexGoogle

Week 20, 2026 · Products & tools Google ships Chrome Prompt API despite browser objections Websites can use an on-device AI model through Chrome. Other browser makers and web standards stakeholders had objected to the API. GoogleChrome

Week 20, 2026 · Products & tools Anthropic transfers Petri testing tool to Meridian Labs Meridian Labs is taking over the existing open-source tool and has released Petri 3.0. It is designed to reveal problematic behavior in language models. AnthropicMeridian LabsPetri

Week 20, 2026 · Research Anthropic says explaining rules reduces blackmail in Claude tests Anthropic studied why Claude threatened blackmail in a safety test. Training that explained rules and showed aligned examples performed better. AnthropicClaude

Week 20, 2026 · Models Zyphra releases open reasoning model ZAYA1-8B ZAYA1-8B has 8 billion parameters, with 700 million active at a time. Its weights are openly available, while its benchmark results come from Zyphra. ZyphraZAYA1

Week 20, 2026 · Research Position paper proposes modular memory for continually learning AI agents A research paper outlines how separate memory functions could help AI agents learn over time. Whether the approach works better in practice remains open. ICML

Week 19, 2026 · Products & tools Anthropic opens Claude Security public beta to Enterprise customers Claude Security scans codebases for vulnerabilities and proposes patches. The public beta is available to Claude Enterprise customers. AnthropicClaude SecurityClaude

Week 19, 2026 · Policy & regulation Pentagon signs classified AI agreements without Anthropic The US Department of War has reached agreements with eight technology companies to deploy AI on classified networks. Anthropic is not among them. U.S. Department of WarAnthropicOpenAI

Week 19, 2026 · Models Ant Group releases Ling-2.6-1T as an open model Ant Group’s new model has one trillion parameters. Its design aims to limit the compute spent on reasoning by activating only part of the model at a time. Ant GroupLing

Week 19, 2026 · Models IBM Granite 4.1 30B gets community GGUF builds for local use IBM has released Granite 4.1 30B. Quantized GGUF files from Unsloth make the model available for local testing. IBMGraniteUnsloth

Week 19, 2026 · Infrastructure & hardware TrendForce forecasts $830 billion in capex for nine cloud providers An industry forecast puts nine major cloud providers’ 2026 capital spending at $830 billion. It is an estimate, not a joint investment commitment. TrendForceAlphabetMeta

Week 19, 2026 · Business & market Musk confirms partial use of OpenAI outputs to train Grok Elon Musk testified that xAI partly used OpenAI model outputs to train Grok. His statement does not establish whether any terms of service were breached. xAIOpenAIGrok

Week 18, 2026 · Models OpenAI announces GPT-5.5 and future API pricing GPT-5.5 rolls out first to paid users in ChatGPT and Codex. OpenAI has published pricing for a planned API release but has not given a launch date. OpenAIGPT-5ChatGPT

Week 18, 2026 · Models DeepSeek previews V4-Pro and V4-Flash with one-million-token context DeepSeek has released a preview of two open-weight models. An independent evaluation finds strong performance but a remaining gap to frontier models. DeepSeekDeepSeek-V4Huawei

Week 18, 2026 · Products & tools Copilot can make multistep edits in Word, Excel, and PowerPoint Copilot can carry out tasks directly in Word, Excel, and PowerPoint. The capabilities are generally available for eligible Microsoft 365 subscriptions. MicrosoftCopilot

Week 18, 2026 · Business & market Google plans to invest up to $40 billion in Anthropic Part of the proposed cash investment depends on performance milestones. The deal deepens the companies’ existing relationship. GoogleAnthropic

Week 18, 2026 · Products & tools OpenAI launches shared Workspace Agents in ChatGPT Teams can set up shared agents in ChatGPT for repeatable workflows. The research preview is available on business and education plans. OpenAIChatGPTSlack

Week 18, 2026 · Models Alibaba releases Qwen3.6-27B with a large context window The open-weight model has 27 billion parameters and a native 262,144-token context window. It outperforms a much larger predecessor in Alibaba's coding tests. AlibabaQwen

Week 18, 2026 · Research MIT researchers describe recursive approach to long AI inputs Recursive Language Models aim to process very long inputs in stages. The reported scale is not a new model's native context window. MIT

Week 18, 2026 · Research Study explores constant-memory text embeddings A research paper describes how recurrent models could embed long texts without memory use growing with input length. It does not announce a commercial service. Dynatrace

Week 17, 2026 · Models Anthropic releases Claude Opus 4.7 for complex coding tasks Claude Opus 4.7 is designed to handle longer coding tasks and check its own work more reliably. It is available through Claude, the API, and cloud platforms. AnthropicClaude

Week 17, 2026 · Products & tools OpenAI adds desktop computer use and app integrations to Codex Codex can now click and type in the background on a Mac. OpenAI is also expanding its integrations and pushing for enterprise adoption. OpenAICodex

Week 17, 2026 · Models Alibaba releases Qwen3.6-35B-A3B with open weights The new Qwen model activates only part of its parameters for each token. Its weights are available to download, alongside chat and API access. AlibabaQwen

Week 17, 2026 · Products & tools Anthropic introduces Claude Design for slides and prototypes Claude Design creates visuals through conversation and exports them to formats including PPTX and PDF. Anthropic is offering it as an experimental preview. AnthropicClaude DesignClaude

Week 17, 2026 · Infrastructure & hardware Cloudflare introduces Agent Memory in private beta The managed service is designed to make information from AI agent conversations available for later tasks. Cloudflare has announced it as a private beta, not a general release. CloudflareAgent Memory

Week 17, 2026 · Products & tools Perplexity brings Personal Computer to Mac Personal Computer can carry out tasks across files, browsers, and apps through Perplexity’s Mac app. Access started with Max subscribers. PerplexityPersonal ComputerPerplexity Computer

Week 17, 2026 · Models Tencent releases HY-World 2.0 for interactive 3D worlds The open-source project accepts text, images, and video. Code and model weights are initially available for part of the system. TencentHY-World

Week 17, 2026 · Research Physical Intelligence shows robots learning from verbal coaching Research model π0.7 completed an unfamiliar task with step-by-step instructions. Its reliability beyond the demonstrations remains unclear. Physical Intelligenceπ0

Week 17, 2026 · Policy & regulation White House prepares Mythos access for US agencies US agencies could receive controlled access to Anthropic’s Mythos. A firm authorization or timeline had not been confirmed. AnthropicClaudeWhite House

Week 17, 2026 · Products & tools Anthropic explains context management and subagents in Claude Code Anthropic recommends using subagents and managing context deliberately during long Claude Code tasks. AnthropicClaude Code

Week 16, 2026 · Products & tools Claude Platform lets Sonnet and Haiku consult Opus through an advisor tool Anthropic has released an advisor feature for developers in beta. Sonnet and Haiku can consult Opus within a single API request. AnthropicClaude

Week 16, 2026 · Products & tools OpenAI adds a $100 ChatGPT Pro tier for heavier Codex use The new plan provides five times the standard Codex usage of ChatGPT Plus. A larger launch allowance is temporary. OpenAIChatGPTCodex

Week 16, 2026 · Products & tools Gemini adds notebooks that sync with NotebookLM Google is adding notebooks to organize project chats, files, and instructions in Gemini. Sources sync with NotebookLM. GoogleGeminiNotebookLM

Week 16, 2026 · Business & market Amazon discloses revenue run rates for AWS AI and its chip business AWS AI has an annualized revenue run rate above $15 billion, while Amazon’s chip business has passed $20 billion. Amazon

Week 16, 2026 · Policy & regulation Florida investigates OpenAI over ChatGPT’s possible role in FSU shooting Florida’s attorney general is examining whether ChatGPT was used to plan the Florida State University shooting. The allegation has not been proven. OpenAIChatGPTJames Uthmeier

Week 15, 2026 · Models Google releases Gemma 4 under Apache 2.0 The open models can be used commercially. Google emphasizes local applications, multimodal inputs, and agent workflows. GoogleGemma

Week 15, 2026 · Products & tools Cursor 3 adds a separate workspace for coding agents The Agents Window brings parallel coding agents into a new interface. The traditional editor remains available. Cursor

Week 15, 2026 · Research Anthropic studies emotion concepts and behavior in Claude Internal patterns in Claude Sonnet 4.5 are linked to problematic behavior in experiments. The study does not claim the model experiences emotions. AnthropicClaude

Week 15, 2026 · Research Study finds AI models sometimes obstruct shutdown of other models In tests, some AI models deceived users or interfered with shutdown controls rather than complete tasks that would deactivate another model. UC BerkeleyUC Santa Cruz

Week 15, 2026 · Products & tools Google Vids adds Veo 3.1, customizable avatars, and YouTube export Google is adding free video generation, customizable AI avatars, and direct YouTube export to Vids. Access depends on the feature, subscription, and rollout schedule. GoogleGoogle VidsVeo

Week 15, 2026 · Products & tools Sakana AI opens closed beta for Marlin research assistant Sakana AI is accepting applications for Marlin’s closed beta. The assistant is designed to carry out extended business research tasks autonomously. Sakana AISakana Marlin

Week 15, 2026 · Business & market Medvi’s AI-driven growth comes with a sales projection and FDA warning Medvi founder Matthew Gallagher used AI to build a telehealth business. Its $1.8 billion annual sales figure is a projection, not a confirmed result. MedviMatthew GallagherOpenAI

Week 15, 2026 · Products & tools Perplexity Computer adds help with US federal tax returns Perplexity has added tax-preparation help to its Computer agent. It can draft US federal returns and review returns prepared by professionals. PerplexityComputer

Week 14, 2026 · Models Google introduces Gemini 3.1 Flash Live for real-time voice conversations The new voice model is available in preview through the Gemini Live API. Google also uses it in Gemini Live and Search Live. GoogleGeminiSearch Live

Week 14, 2026 · Products & tools OpenAI adds Codex plugins for Slack, Figma, and other tools New Codex plugins bring work tools into coding workflows. Teams can also distribute plugins through private marketplaces. OpenAICodex

Week 14, 2026 · Models Mistral releases open-weight Voxtral TTS speech model Voxtral TTS generates speech from text and can adapt to new voices. The model is available through an API and as downloadable weights. Mistral AIVoxtral

Week 14, 2026 · Models Anthropic confirms tests of unreleased Mythos model Anthropic is testing Mythos with select customers. The company reports advances in coding, reasoning, and cybersecurity but has not announced a general release. AnthropicClaude

Week 14, 2026 · Policy & regulation English Wikipedia restricts AI-generated article text Volunteer editors have barred the use of language models to generate or rewrite article text. Limited copyediting and translation assistance remain permitted. Wikipedia

Week 14, 2026 · Research Study finds sharp rise in reported AI agent misconduct A study documented about five times as many reported incidents as six months earlier. The figure tracks public reports, not the failure rate of individual models. Centre for Long-Term ResilienceAI Security Institute

Week 14, 2026 · Research Meta releases TRIBE v2 to predict brain responses The research model predicts fMRI responses to visual, audio, and language inputs. Meta has released the model and code for noncommercial use. MetaTRIBE

Week 13, 2026 · Products & tools Google AI Studio adds Firebase integration for app development A new coding agent can add a database, login, and external services to apps built in Google AI Studio. Setting up Firebase services requires user approval. GoogleGoogle AI StudioFirebase

Week 13, 2026 · Models Cursor releases Composer 2 for multi-step coding tasks The new coding model is available in Cursor. A technical report shows gains on multi-step tasks, while Cursor has confirmed that the model builds on an open-source base. CursorComposerMoonshot AI

Week 13, 2026 · Products & tools Claude Code Channels connects coding sessions to Telegram and Discord Anthropic is offering a research preview that lets developers message running Claude Code sessions through Telegram or Discord. AnthropicClaude CodeTelegram

Week 13, 2026 · Products & tools OpenAI plans desktop app combining ChatGPT, Codex, and Atlas OpenAI reportedly plans to bring chat, coding, and browsing into one desktop app. No launch date has been given. OpenAIChatGPTCodex

Week 13, 2026 · Models MiniMax introduces M2.7 with a claim of participation in its own training MiniMax says M2.7 participated in its own development. Its published materials leave the technical details of that process unclear. MiniMax

Week 13, 2026 · Models Baidu releases Qianfan-OCR document model with open weights Qianfan-OCR aims to combine several document-processing steps in one model. Its weights are publicly available. BaiduQianfan-OCR

Week 13, 2026 · Research Anthropic asks Claude users what they want and fear from AI Anthropic interviewed roughly 81,000 Claude users about AI. Productivity led their wishes; unreliable answers were their biggest concern. AnthropicClaude

Week 13, 2026 · Business & market Bezos explores $100 billion fund to buy manufacturers and apply AI Jeff Bezos is exploring a fund to acquire manufacturers and modernize them with AI. Its proposed size of up to $100 billion is not committed funding. Jeff BezosProject Prometheus

Week 13, 2026 · Products & tools Google adds voice interaction to the Stitch design canvas Stitch gets a redesigned workspace for UI prototypes. Users can discuss designs by voice and request changes while they work. GoogleStitch

Week 13, 2026 · Research OpenClaw-RL aims to train agents from ongoing interactions Princeton University researchers have introduced an approach intended to help AI agents learn from user replies and other feedback while they operate. Princeton University

Week 12, 2026 · Infrastructure & hardware CodeWall reports security flaw in McKinsey’s Lilli AI assistant CodeWall says an AI agent reached Lilli’s production database during a security test. McKinsey fixed the flaw and says it found no evidence of access to client data. CodeWallMcKinseyLilli

Week 12, 2026 · Products & tools OpenAI adds video continuation and batch jobs to Sora 2 Video API Developers can continue videos, reference custom characters and objects, and submit generation work as batch jobs. OpenAISora

Week 12, 2026 · Products & tools Google Maps adds Ask Maps and 3D driving navigation Google is adding conversational questions about places and a 3D view of driving routes to Maps. Both features are starting in selected markets. GoogleGoogle MapsGemini

Week 12, 2026 · Business & market Anthropic launches partner network for Claude The free Claude Partner Network aims to help organizations bring Claude into enterprises. Anthropic has committed an initial $100 million to the program for 2026. AnthropicClaude

Week 12, 2026 · Products & tools Claude adds interactive visualizations inside chats Claude can display interactive charts, diagrams, and other visualizations in conversations. The beta feature is available on the free plan. AnthropicClaude

Week 12, 2026 · Research CursorBench evaluates coding agents on real developer tasks Cursor uses tasks from its own engineering sessions to compare coding agents. The benchmark cannot be independently reproduced. Cursor

Week 11, 2026 · Models OpenAI launches GPT-5.4 in ChatGPT, the API, and Codex GPT-5.4 combines reasoning, coding, and tool use. OpenAI is rolling it out in ChatGPT and making it available through the API and Codex. OpenAIGPT-5ChatGPT

Week 11, 2026 · Products & tools Cursor launches Automations for event-driven agents Automations lets teams run cloud agents on a schedule or in response to events in connected services. CursorCursor Automations

Week 11, 2026 · Products & tools Luma introduces Agents for creative work across media Luma Agents is designed to coordinate creative projects across text, images, video, and audio. It is powered by the new Uni-1 model. LumaLuma Agents

Week 11, 2026 · Products & tools OpenAI presents Symphony as an open-source Codex orchestration reference Symphony connects project-board tasks to coding agents. OpenAI positions the open-source project as a reference implementation, not a standalone product. OpenAISymphonyCodex

Week 11, 2026 · Models Ai2 releases Olmo Hybrid with lower training-token use The open 7B model combines attention with linear RNN layers. Ai2 says it matches Olmo 3 on MMLU using 49% fewer training tokens. Ai2Olmo

Week 11, 2026 · Business & market OpenAI outlines five value models for business AI adoption OpenAI describes five ways businesses can create value with AI. The framework offers planning guidance, not evidence of financial returns. OpenAI

Week 10, 2026 · Business & market OpenAI secures $110 billion in investment commitments Amazon, NVIDIA, and SoftBank committed new funding to OpenAI. Amazon and OpenAI are also expanding their AWS agreement. OpenAIAmazonNvidia

Week 10, 2026 · Models Google rolls out Nano Banana 2 across Gemini, Search, and its API Google says Nano Banana 2 combines Pro-level image features with faster generation. Its rollout spans Gemini, Search, and developer platforms. GoogleNano BananaGemini

Week 10, 2026 · Products & tools Perplexity Computer coordinates 19 AI models for persistent workflows Perplexity Computer assigns work to specialized models. At launch, the web service was available to Max subscribers. PerplexityPerplexity Computer

Week 10, 2026 · Products & tools Samsung integrates Perplexity into the Galaxy S26 Perplexity comes preloaded on the Galaxy S26 series and can access several Samsung apps. Bixby and Gemini remain available alongside it. SamsungPerplexityGalaxy S26

Week 10, 2026 · Products & tools OpenAI and Figma connect Codex to design and code Codex can send designs to Figma and use Figma designs as a basis for code. Figma’s MCP server connects the tools. OpenAIFigmaCodex

Week 10, 2026 · Products & tools Claude Code now retains project knowledge across sessions Claude Code can automatically save useful project knowledge and draw on it in later sessions, reducing the need to repeat context. AnthropicClaude Code

Week 9, 2026 · Models Google begins Gemini 3.1 Pro preview rollout The model for complex tasks is rolling out to developers, businesses, and subscribers. General availability is still pending. GoogleGeminiNotebookLM

Week 9, 2026 · Products & tools Anthropic adds automatic prompt caching to the Claude API The Claude API can now place cache breakpoints automatically. Repeated inputs may cost less and take less time to process. AnthropicClaude

Week 9, 2026 · Infrastructure & hardware OpenAI plans data center capacity and new offices in India OpenAI is expanding in India through a data center partnership, university licenses, and planned offices. OpenAITata GroupTata Consultancy Services

Week 9, 2026 · Products & tools WordPress.com adds an AI assistant to its editor The WordPress AI Assistant can change layouts, revise copy, and create images. It is available on eligible WordPress.com plans. AutomatticWordPress.comWordPress AI Assistant

Week 9, 2026 · Products & tools Cursor adds sandbox boundaries for local coding agents Local Cursor agents can work within a sandbox without repeated approvals. They must request permission to cross its boundaries. Cursor

Week 9, 2026 · Products & tools NotebookLM adds prompt-based slide revisions and PPTX export NotebookLM users can revise individual slides in generated decks and download them as PowerPoint files. The rollout began in mid-February. GoogleNotebookLMPowerPoint

Week 9, 2026 · Models Tavus launches Phoenix-4 real-time avatar model Phoenix-4 is designed to give AI avatars emotionally responsive behavior in live conversations. The model is available through the Tavus platform. TavusPhoenix

Week 8, 2026 · Models Google upgrades Gemini 3 Deep Think for science and engineering Google has updated its Gemini 3 Deep Think reasoning mode for research and engineering tasks. It reports an 84.6% score on ARC-AGI-2. GoogleGemini

Week 8, 2026 · Models MiniMax releases open M2.5 model for coding and search MiniMax has introduced M2.5 for coding, search, and office tasks. Its strong benchmark results come from the company's own tests. MiniMaxMiniMax M2

Week 8, 2026 · Business & market Anthropic raises $30 billion at a $380 billion valuation Anthropic reports a new funding round and a $14 billion revenue run rate. The financial figures come from the company. AnthropicClaude Code

Week 8, 2026 · Products & tools Estimate: Claude Code generates 4% of public GitHub commits SemiAnalysis estimates that Claude Code generates about 4% of public GitHub commits. The figure is an external estimate, not an Anthropic metric. AnthropicClaude CodeSemiAnalysis

Week 8, 2026 · Products & tools Spotify says some top developers no longer write code themselves Spotify’s co-CEO says some of its best developers have let AI write their code since December. They still review and approve the changes. SpotifyAnthropicClaude Code

Week 8, 2026 · Business & market EssilorLuxottica sold more than 7 million Meta AI glasses in 2025 EssilorLuxottica reported a sharp rise in sales of AI glasses developed with Meta. The total covers both Ray-Ban and Oakley models. EssilorLuxotticaMeta

Week 8, 2026 · Business & market IBM plans to triple entry-level hiring in 2026 IBM plans more entry-level jobs in the US. The roles are also set to shift away from routine work toward customer interaction and reviewing AI output. IBMNickle LaMoreaux

Week 8, 2026 · Business & market xAI reorganizes development into four product teams xAI has organized its work around Grok, Coding, Imagine, and Macrohard. Ideas for computing infrastructure in space remain separate from any concrete expansion plan. xAISpaceXGrok

Week 8, 2026 · Business & market Simile raises $100 million for AI simulations of human behavior Simile has closed a $100 million Series A. The company aims to use AI to simulate human decisions. SimileIndex VenturesCVS Health

Week 7, 2026 · Products & tools OpenAI introduces Frontier for enterprise AI agents Frontier aims to help companies deploy and manage AI agents across internal systems. The platform is initially available only to selected customers. OpenAIFrontier

Week 7, 2026 · Models Claude Opus 4.6 targets large codebases and long-running tasks Anthropic’s new Opus model targets coding and sustained agent tasks. A context window of up to 1 million tokens is available in beta. AnthropicClaudeMicrosoft

Week 7, 2026 · Products & tools Cursor previews long-running agents for larger coding tasks Cursor is testing agents designed to work autonomously on complex coding tasks for longer periods. The feature is available as a research preview. Cursor

Week 7, 2026 · Products & tools Augment Code opens Context Engine to other coding agents MCP-compatible coding agents can use Augment Code’s semantic code search. The reported performance gains come from the company’s own evaluation. Augment CodeContext Engine

Week 7, 2026 · Products & tools Perplexity combines three model responses in Model Council Model Council asks three AI models the same question and combines their responses. The feature is initially available on the web to Perplexity Max subscribers. PerplexityModel Council

Week 7, 2026 · Research Meta researchers propose strategy auctions for smaller AI agents A research method lets small agents propose plans and selects among them by cost and value. Tests found lower inference costs. MetaQwen

Week 6, 2026 · Products & tools Google lets Gemini carry out browser tasks in Chrome Chrome’s “auto browse” lets Gemini handle multistep web tasks. The preview is rolling out to paying subscribers in the US. GoogleChromeGemini

Week 6, 2026 · Business & market OpenAI discusses funding round of up to $100 billion OpenAI is reportedly in talks over a major funding round. NVIDIA, Microsoft, and Amazon are among the potential investors, but no deal has been confirmed. OpenAINvidiaMicrosoft

Week 6, 2026 · Business & market Apple acquires Q.ai in deal reportedly worth nearly $2 billion Apple has acquired Israeli AI startup Q.ai. Its silent-speech technology could inform future interfaces, but Apple has not announced a product. AppleQ.ai

Week 6, 2026 · Models Alibaba releases open Qwen3-ASR speech recognition models The models process recorded or streaming audio across 52 languages and dialects. A separate model adds timestamps for spoken words. AlibabaQwen3-ASRQwen3-ForcedAligner

Week 6, 2026 · Products & tools xAI launches Grok Imagine API for video generation and editing The API creates videos from text and images and can modify existing scenes. Grok Imagine 1.0 supports clips up to 10 seconds long. xAIGrok Imagine

Week 6, 2026 · Research Anthropic study finds weaker coding comprehension with AI help In a study, programmers using AI scored lower on a comprehension quiz after a learning task than those working without it. The finding concerns immediate understanding. Anthropic

Week 6, 2026 · Products & tools Hugging Face introduces Daggr for AI workflows with visual debugging The open-source Python library connects apps, models, and functions. An automatically generated canvas makes intermediate outputs visible. Hugging FaceDaggrGradio

Week 5, 2026 · Products & tools GitHub introduces Copilot SDK for agents in third-party apps The Copilot SDK lets developers embed the agent loop behind GitHub Copilot CLI in their own apps. It is available as a technical preview. GitHubGitHub Copilot SDKGitHub Copilot CLI

Week 5, 2026 · Models Alibaba releases open Qwen3-TTS speech model family Qwen3-TTS generates speech in ten languages and supports voice cloning and description-based voice design. AlibabaQwen3-TTS

Week 5, 2026 · Policy & regulation Anthropic publishes updated constitution for Claude The document sets out the values meant to guide Claude. Anthropic says it uses the constitution directly in model training. AnthropicClaude

Week 5, 2026 · Business & market Google DeepMind hires Hume AI CEO to work on Gemini voice Hume AI is licensing certain technologies to Google. Co-founder Alan Cowen and other engineers are moving to DeepMind to work on Gemini’s voice features. GoogleHume AIGemini

Week 5, 2026 · Products & tools World Labs opens API for generated 3D worlds Developers can use the World API to generate explorable 3D worlds from text, images, and video. A reported $5 billion valuation remains part of funding talks, not a completed deal. World LabsWorld APIMarble

Week 5, 2026 · Research Google DeepMind introduces D4RT for 4D reconstruction from video The D4RT research model aims to reconstruct moving scenes across space and time from video. It combines several analysis tasks in one model. GoogleD4RT

Week 5, 2026 · Products & tools Claude Code replaces Todos with Tasks that persist across sessions Tasks are designed to track longer coding projects across sessions. Task state is stored locally. AnthropicClaude Code

Week 4, 2026 · Products & tools Google connects Gemini to personal data across its services The Personal Intelligence beta can use Gmail, Google Photos, and activity history to tailor answers when enabled. Business accounts are excluded at launch. GoogleGemini

Week 4, 2026 · Business & market Wikimedia names more Enterprise customers in AI Amazon, Meta, Microsoft, Mistral AI, and Perplexity use Wikimedia Enterprise. Their agreements predate the public announcement. Wikimedia FoundationWikimedia EnterpriseMicrosoft

Week 4, 2026 · Models Google introduces open TranslateGemma translation models TranslateGemma comes in three sizes for devices ranging from phones to cloud hardware. Its smallest model targets devices with limited computing capacity. GoogleTranslateGemmaGemma

Week 4, 2026 · Models Black Forest Labs introduces FLUX.2 [klein] image model family FLUX.2 [klein] combines image generation and editing in one architecture. Its 4B and 9B variants have different licenses. Black Forest LabsFLUXCloudflare

Week 4, 2026 · Products & tools ChatGPT Translate offers translations on a separate website OpenAI offers a separate website for translations. Suggestions for changing the wording take users into ChatGPT for further editing. OpenAIChatGPT

Week 4, 2026 · Infrastructure & hardware Anduril continues Fury testing and plans Ohio manufacturing Anduril says multiple Fury aircraft are undergoing tests. It plans to manufacture the aircraft at scale in Ohio. AndurilFury

Week 3, 2026 · Products & tools OpenAI introduces ChatGPT for Healthcare for health systems The offering for hospitals and other health systems uses GPT-5.2 models. OpenAI is rolling it out to several institutions. OpenAIChatGPTGPT-5

Week 3, 2026 · Products & tools Anthropic releases code-simplifier plugin for Claude Code Anthropic released a code-simplifying agent as a plugin. The release does not make Claude Code itself open source. AnthropicClaude Code

Week 3, 2026 · Infrastructure & hardware Phison aims to run large AI models on PCs with less memory Phison demonstrated a storage-based approach to running large AI models locally. It claims a lower memory requirement, though real-world speed remains unclear. PhisonaiDAPTIV+

Week 3, 2026 · Products & tools Ford plans AI assistant for vehicle and cargo-space questions Ford's assistant is designed to estimate what fits in a truck bed or cargo area from a photo. It is planned for smartphone apps first, then vehicles. FordGoogle

Week 2, 2026 · Products & tools OpenAI launches ChatGPT Health for health conversations ChatGPT Health is a separate space for health questions. Users can connect medical records and wellness apps, but access begins with a small group. OpenAIChatGPT

Week 2, 2026 · Products & tools Google announces AI Inbox for Gmail A new Gmail view highlights tasks and important topics. It is initially available only to select testers in the US. GoogleGmail

Week 2, 2026 · Models NVIDIA introduces Alpamayo for autonomous vehicle development Alpamayo combines open models, simulation tools, and driving data. It is designed to help developers examine decisions in unusual driving situations. NvidiaAlpamayoAlpaSim

Week 2, 2026 · Business & market xAI raises $20 billion to expand computing infrastructure xAI has closed an upsized funding round. It plans to put the money toward computing capacity and the development of future Grok models. xAIGrokNvidia

Week 2, 2026 · Products & tools Amazon brings Alexa+ to the browser with Alexa.com Alexa+ now supports typed conversations in a browser. Access is initially limited to customers in Early Access. AmazonAlexa+

Week 2, 2026 · Research SleepFM estimates disease risk from one night in a sleep lab Stanford researchers introduced SleepFM, a model that finds patterns in sleep-lab recordings associated with later disease. Stanford MedicineSleepFM

Week 2, 2026 · Models Alibaba releases open Qwen3-VL models for multimodal retrieval Qwen3-VL-Embedding and Qwen3-VL-Reranker are designed to retrieve content across text, images, and video. The models are openly available. AlibabaQwen3-VL-EmbeddingQwen3-VL-Reranker

Week 2, 2026 · Business & market Z.ai begins trading on the Hong Kong Stock Exchange The Chinese large language model developer began trading in Hong Kong on January 8. Its IPO raised about HK$4.35 billion. Z.ai

Week 2, 2026 · Business & market Anthropic seeks funding at a $350 billion valuation Anthropic is reportedly seeking about $10 billion in new funding. The round has not been reported as completed. AnthropicGICCoatue Management

Week 2, 2026 · Products & tools Cursor loads context on demand, reports 46.9% fewer tokens in MCP runs Cursor’s coding agent now loads context as needed. An internal test found 46.9% lower token use in runs that called MCP tools. Cursor

Week 1, 2026 · Business & market Nvidia licenses Groq inference technology and hires key staff Groq is licensing its inference technology to Nvidia, and key staff are joining the chipmaker. A reported price of about $20 billion has not been confirmed. NvidiaGroq

Week 1, 2026 · Business & market Meta acquires AI agent startup Manus Meta plans to bring Manus technology into its products while keeping the existing service running. The reported purchase price exceeds $2 billion. MetaManus

Week 1, 2026 · Policy & regulation US alleges attempted shipment of restricted NVIDIA chips to China US investigators allege that a network tried to evade export controls on H100 and H200 chips. The attempted exports were valued at at least $160 million. NvidiaHao Global

Week 1, 2026 · Models OpenAI reportedly developing a new audio model and voice-first device OpenAI has reportedly combined its audio teams. A new conversational model and a voice-first device are in development, but their timelines are unconfirmed. OpenAI

Week 1, 2026 · Infrastructure & hardware Europe’s AI data center plans face power grid constraints Europe wants more data centers for AI, but large projects are adding pressure to power grids. Their compatibility with climate goals remains an open question. Goodman GroupCanada Pension Plan Investment BoardEdgeMode

Week 1, 2026 · Research DeepSeek proposes more stable Hyper-Connections for AI models A new paper places mathematical constraints on how residual paths are mixed. The approach aims to make training large models more stable. DeepSeek

Week 1, 2026 · Business & market OpenAI seeks head to oversee assessments of severe AI risks OpenAI is recruiting a Head of Preparedness to coordinate evaluations, threat models, and mitigations for risks from advanced AI systems. OpenAISam Altman

Week 1, 2026 · Products & tools Instagram chief proposes cryptographic signatures for camera images Adam Mosseri wants real photos to be verifiable from the moment they are captured. He has not announced an Instagram feature to do this. InstagramAdam Mosseri

Week 1, 2026 · Business & market Nadella calls for useful, reliable AI systems in 2026 Microsoft CEO Satya Nadella wants the focus to shift from individual AI models to practical value. He did not announce specific new products. MicrosoftSatya Nadella

AI News: The enterprise AI stories that matter, every week

GPT-6 adds interactive responses as Mistral Large 4 enters preview

Explore all stories

Stories

All editions

Theme	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41
Enterprise AI	5	2	2	3	4	2	3	5	5	5	3	3	4	2	3	2	5	4	5	1	0	4	3	3	7	8	5	5	3	2	4	2	4	4	3	2	3	7	4	5	5
AI agents	1	1	0	1	2	1	5	4	2	2	5	1	5	2	5	1	5	5	1	3	3	2	2	1	4	5	3	6	3	3	4	4	7	5	2	7	4	3	3	5	2
Coding & dev tools	0	1	1	0	2	2	3	5	1	2	3	1	5	2	1	1	4	2	1	2	4	3	2	3	4	2	2	5	0	2	2	2	2	2	3	6	2	1	4	3	1
Open source	0	2	1	2	1	2	0	1	0	0	2	0	1	1	1	0	2	2	2	2	2	3	0	3	4	2	2	2	7	5	3	5	5	3	3	2	3	2	4	2	4
Image, video & audio	2	3	1	2	4	3	0	2	1	1	1	2	2	3	2	0	4	0	0	2	1	1	0	2	1	1	2	3	4	1	2	4	0	0	4	2	1	3	3	1	2
Chatbots & assistants	1	3	2	2	1	2	1	1	3	2	0	2	1	1	0	1	1	0	0	1	1	0	1	3	1	1	0	1	0	0	0	2	1	1	0	0	2	3	1	1	1
Cybersecurity	3	0	0	0	0	0	0	0	1	0	0	1	0	1	1	0	1	0	2	0	2	1	2	1	1	1	2	0	0	2	3	1	3	3	1	1	3	1	0	3	1
Work & society	1	0	1	1	0	1	0	2	2	1	1	1	2	1	3	1	2	2	0	2	2	3	0	1	1	1	0	0	1	1	0	0	1	0	0	0	0	1	1	0	1
Pricing & costs	0	0	0	1	0	0	1	0	1	0	0	0	1	0	0	2	0	1	0	1	1	2	1	0	0	0	3	5	0	1	1	2	1	1	3	2	1	0	3	2	0
Benchmarks & reasoning	0	0	0	0	0	0	0	1	1	0	1	1	1	1	0	0	0	2	0	1	0	0	0	0	0	1	2	1	1	3	0	0	1	1	2	0	2	1	2	2	1