Meta releases Muse Spark Open as Anthropic previews Claude Duet

Anthropic introduced Claude Duet on October 21, 2026, as a research preview.¹ Powered by Claude Opus 5.5, the assistant continues working in a virtual machine with its own browser, so a user’s device does not need to stay on.¹ Pro and Max subscribers in the US can join a waitlist for access.² Duet prepares morning briefs, monitors connected services, and alerts users when a decision is needed.¹

Anthropic lists scheduling, forecasting, and building dashboards and web applications among its potential uses.¹ A Stripe integration also lets Duet make approved purchases with a virtual card.³ The product is aimed at workflows that span multiple applications rather than single questions. Sensitive actions, including sending messages, making payments, and deleting files, require approval.¹² Whether those workflows prove reliable in everyday use remains an open question during the limited preview.

Meta released Muse Spark Open on October 15, 2026, and made its model weights available for download.¹ The open-weight model has 816 billion parameters, with about 42 billion active per token.¹² It accepts text, images, audio, and video as input and produces text.² Meta designed it for coding, tool use, and multistep agent tasks.¹² It is also available through several API providers.¹

The model follows Muse Glimmer in Meta’s Muse series.² Meta is working with Apple to optimize local inference using MLX.¹ That setup calls for specialized hardware, not an ordinary Mac.¹³ The Muse Model License also restricts certain commercial uses, including resale as a hosted service.⁴ Meta’s published benchmark scores reflect its own tests; they do not independently establish performance in real-world workflows.⁴

OpenAI and Google Cloud announced on October 21 that OpenAI models and agentic Codex workflows are coming to Google Cloud. The models will be added to Model Garden, allowing enterprises to use Google Cloud infrastructure, access controls, and data governance capabilities. Access is starting in private preview for selected enterprise customers, with purchases planned through Google Cloud Marketplace. General availability is expected in the coming weeks, but no firm date was given.¹²

OpenAI models are also planned for Google Antigravity and Gemini Agent; those integrations are not yet complete.¹² The agreement broadens model choice for Google Cloud customers and gives OpenAI another enterprise distribution channel. OpenAI says Microsoft remains its primary cloud partner.¹ Neither company has published prices, supported regions, or service-level commitments.¹²

Two critical PHP vulnerabilities could allow unauthenticated remote code execution under certain conditions.¹² Wiz used OpenAI’s GPT-5.6 Sol, Google’s Gemini 4 Argon, and its Atlas agent to find and test the flaws.² Security firm Calif had reportedly found them in August during an OpenAI-funded code audit.² PHP released security updates on October 21.¹

The flaws involve memory management and can be chained together.¹³ Palo Alto Networks identified more than 22.6 million publicly exposed instances that might be vulnerable; that figure is not a count of confirmed attacks.² Exploitation attempts have already been reported.¹² For operators, the immediate priority is updating affected installations and checking internet-facing systems. The scale of confirmed compromises remains unclear.

DeepSeek made DeepInfer open source on October 20, 2026. The company says the framework powers its production inference services.¹² The initial public release targets offline inference for DeepSeek-V4.1-Flash. It includes optimized kernels, distributed inference, and speculative decoding.¹³ DeepSeek also released deepseek-recipe, a component intended to reproduce the token-processing behavior of its official API in community deployments.²

DeepSeek aims to reduce compute costs and energy use, but its announcement provides no measurements or independent tests showing the size of any savings.² The published code is not a complete online-serving system: an analysis of the release found no KV-cache management, prefill/decode disaggregation, or data-parallel load balancing.⁴ Companies therefore need to assess both the performance gains and the integration work against their own workloads.

DeepSeek updated its V4.1 Flash model on October 17, 2026. DeepSeek-V4.1-Flash-1017 is available under the existing deepseek-flash API name, so current integrations need no code changes, according to the API documentation.¹ The model weights were released under the MIT license.² DeepSeek also opened public Search and Fetch APIs: Search returns web results, while Fetch retrieves individual pages.³⁴

The endpoints can help coding agents look up information on the web. They are not generally free: standalone Search and Fetch calls each cost $0.75 per 1,000 successful requests.³ Only server-side calls through DeepSeek Harness, Claude Code, and Codex are temporarily free for DeepSeek API users.³ Teams building agents should distinguish that limited offer from ordinarily billed API calls when estimating operating costs.

NVIDIA released Nemotron 3.5 Super VL on October 15 as an open-weight multimodal model for AI agents. It handles text, images, documents, and video, and is designed to combine reasoning, instruction following, and tool use in one model.¹² NVIDIA makes it available through Hugging Face, OpenRouter, ModelScope, and NVIDIA NIM, with deployment options including workstations and data centers.¹

The model targets workflows in which agents examine material such as tables, charts, or web pages.¹ NVIDIA claims up to five times the throughput of other open models in its size class. That comparison comes from the company and is not independently verified by the cited sources.¹ Companies considering it for production should test it against their own documents and videos.

Inception introduced Mercury 3 on October 19, 2026. The language model targets applications where response time matters and is available through the Inception API, Baseten, OpenRouter, and Vercel.¹ Instead of producing an answer one token at a time, it drafts multiple parts in parallel and refines them over several passes.² Inception reports a speed of 1,068 tokens per second on NVIDIA GPUs.²

The launch puts response latency alongside quality as a factor in choosing a language model. In Inception's comparison, Mercury 3 led eight of 12 tests against GPT-6 Luna and Claude Haiku 5.5, while the competing models won the others.² The speed and benchmark figures are company-reported, not a guarantee of performance in a particular deployment. Hardware, workload, and API provider can affect the results teams see.²

Stanford’s Hazy Research group has introduced Next-Concept Modeling (NCM). Its project post was updated on October 19.¹ The approach first uses an autoencoder to compress text into a hierarchy of discrete codes. A transformer then generates that hierarchy from coarse to fine, predicting the codes at each level in parallel.¹ The team has released code and pretrained checkpoints.²

The results point to a potential route to faster text generation, not a general speed advantage over today’s large language models. The reported comparisons cover selected tasks and diffusion models of similar scale; both speed and text quality depend on the test.¹ For companies, NCM remains a research approach that needs evaluation on their own workloads and hardware before its production performance can be assessed.

OpenAI published results for COBOL program migration with GPT-6.1 Sol on October 19, 2026.¹ At low reasoning effort, the model passed 86.6% of the test tasks, compared with 5.4% for GPT-5.6 Sol under the same conditions.¹ The task was to translate programs into other languages while preserving their observable behavior.¹ The publication presents new test results for a model that had already been introduced.²¹

OpenAI checked the generated programs with hidden tests requiring byte-for-byte matching outputs. A migration counted as successful only if it passed every test.¹ The result measures performance on these tasks, not whether existing COBOL systems can be migrated safely in production.¹ Production migrations still require expert review and testing against business rules and dependencies.¹ The results come from OpenAI and do not establish a direct comparison with competing models.¹