GPT-5.5 rolls out as DeepSeek previews V4-Pro and V4-Flash

OpenAI announced GPT-5.5 on April 23, 2026.¹ The model rolled out first to paid users in ChatGPT and Codex, with developer access through the API expected later.¹² OpenAI did not give a specific date for API availability.¹ It also announced a higher-end variant called GPT-5.5 Pro.¹

OpenAI emphasizes stronger reasoning, multistep coding, and tool use.¹ It says the model can plan tasks, check its own work, and operate with limited human direction.¹ For companies, the announced token prices make preliminary cost estimates possible, but production API testing must wait for access.¹ The published benchmark scores offer a starting point for evaluating the model, not a substitute for testing it on a company’s own workflows.¹

DeepSeek released previews of DeepSeek-V4-Pro and DeepSeek-V4-Flash on April 24, 2026.¹ Both models have open weights and a one-million-token context window.¹ The API was available at launch, and DeepSeek also offered the models through chat.deepseek.com and for local download.¹ The company designed the V4 series for agents and use with tools such as Claude Code and OpenClaw.¹

V4-Pro uses a mixture-of-experts architecture, while Flash is the smaller variant.¹ The models are adapted for Huawei Ascend chips.² An independent evaluation puts the performance claims in perspective: the US-based CAISI rated V4-Pro the most capable Chinese model it had tested, but estimated that it still trailed frontier models by about eight months.³ Its results were more cautious than DeepSeek’s own evaluations.³

Microsoft made Copilot’s agentic capabilities in Word, Excel, and PowerPoint generally available on April 22, 2026. They are the default experience for Microsoft 365 Copilot and Microsoft 365 Premium subscribers and also extend to Microsoft 365 Personal and Family plans. Copilot can perform multistep tasks directly in files: rewriting text in Word, changing workbooks in Excel, and creating or updating presentations in PowerPoint.¹

This moves Copilot beyond suggestions in a chat window to changes inside the applications. Microsoft says users remain in control.¹ For businesses, the practical scope includes editing formulas, tables, and visuals in Excel and following company templates in PowerPoint.¹ The announcement does not provide benchmark results quantifying time savings or output quality.¹

Google announced on April 24, 2026, that it plans to invest up to $40 billion in Anthropic. The companies are deepening an existing partnership, and the funding is intended to support Anthropic’s expansion of computing capacity.¹² The full amount is not guaranteed: part of it depends on Anthropic meeting agreed performance milestones.²

For Google, the deal adds an investment in an outside model developer alongside its own Gemini models.³ Reuters and Bloomberg reported that the stated investment amount covers cash, not the value of TPU capacity.²⁴ The financing and the companies’ computing arrangements should therefore be assessed separately. The confirmed investment terms do not establish how much additional TPU capacity Anthropic will receive or when it would become available.²⁴

OpenAI introduced Workspace Agents in ChatGPT as a research preview on April 22, 2026. Teams can create shared agents for repeatable, long-running workflows across connected organizational tools. The feature is available for ChatGPT Business, Enterprise, Edu, and Teachers. Slack is among the collaboration services OpenAI mentions, but the agents are not limited to it.¹²

OpenAI positions Workspace Agents as an evolution of GPTs, aimed at shared processes rather than one-off personal assistance.¹² Credit-based billing is planned after an initial free period, but OpenAI has not published a per-credit price.¹ Before a broad rollout, companies will need to decide what data agents may access across connected services and how to track usage once billing starts.

Alibaba released Qwen3.6-27B on April 22 as an open-weight model under the Apache 2.0 license. The dense model has 27 billion parameters and a native 262,144-token context window. Its weights are available on Hugging Face. Qwen3.6-27B supports modes with and without explicit reasoning steps.¹²

On coding tasks, Qwen reports stronger results than for Qwen3.5-397B-A17B, a predecessor with 397 billion total parameters and 17 billion active parameters. Qwen says the new model scores 77.2 versus 76.2 on SWE-bench Verified.¹ Those figures come from Qwen's published tests and do not replace evaluation on a company's own tasks. For longer conversations, the model also offers an option to retain reasoning from earlier messages.³²

MIT researchers describe Recursive Language Models (RLMs) in a research paper discussed in April 2026. The approach is designed to process very long inputs.¹² Rather than placing all the text in one model call, it treats the input as an external environment and uses recursive calls for smaller tasks.² This is an inference-time method, not a new language model with a comparably large native context window.² No commercial launch or price has been announced.²

The reported scale refers to how much text the procedure can handle, not the capacity of any single model call.³² That distinction matters for work with extensive documents, where a model may need to locate relevant information or tackle a task in stages.² The token count alone does not establish how reliable or cost-effective the approach will be for a particular use case.

A paper posted on April 20, 2026, describes a text-embedding method whose computational cost grows linearly with input length.¹ Its vertically chunked inference strategy is designed to keep memory use independent of text length once the input exceeds the chunk size.¹ The researchers examine recurrent language models and use fine-tuned Mamba2 models as general-purpose text embedders.¹

The paper reports results competitive with transformer-based embedding models across multiple benchmarks.¹ That approach could matter for applications that index or search very long documents without increasing memory requirements for every additional passage. The quality claims come from the research itself and do not establish that the method is generally superior to transformers.¹² The available information does not announce a commercial service.¹