This Week in AI — Week of September 21, 2026

This Week in AI is our weekly digest of the stories that actually matter in AI — sourced first party wherever possible, with the original link attached. Here’s the week of September 21, 2026.

🔝 Top story

Claude Opus 5.5, and Anthropic’s first release since it called for pacing

Anthropic shipped Claude Opus 5.5 on September 22, the first model in the 5.5 family and the company’s first release since Dario Amodei’s essay urging the industry to slow down. The framing is efficiency rather than a capability ceiling: it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Early testers reported one agent completing a 680,000-line code migration in under a day, and Anthropic published its strongest alignment scores to date alongside it.

🔗 https://www.anthropic.com/claude-opus-5-5

So the lab that asked everyone else to slow down shipped a flagship the same month. Whether that counts as keeping the promise depends on what the benchmarks do next.

⚡ Models

GPT-6 Sol and Luna bring the Astra generation down to half price

OpenAI added GPT-6 Sol and GPT-6 Luna on September 22, trained with similar methods to GPT-6 Astra to carry that generation’s advances into faster, cheaper models. Sol drops to $2 input and $10 output per million tokens, from $4 and $20, and Luna to $0.10 and $0.50. Both started landing in ChatGPT Work and Codex, which is the first time this generation has been available outside the top tier.

🔗 https://openai.com/index/introducing-gpt-6-sol-and-luna/

MiMo-V2.6-Pro tops the open weight leaderboard

Xiaomi fully open sourced the MiMo-V2.6 series on September 22: the flagship MiMo-V2.6-Pro and the efficient MiMo-V2.6-Flash, both natively omnimodal, plus a Pro-UltraSpeed mode generating at up to 20x speed. Pro scores 46.32 on the Artificial Analysis Intelligence Index, the highest of any open weight model, ahead of Kimi K3 and Qwen3.8 Max, with the RL training code released alongside the weights.

🔗 https://mimo.xiaomi.com/mimo-v2-6/article

FLUX 3 Action turns a camera frame into the next two seconds of action

Black Forest Labs released FLUX 3 Action, a 7B open weights world action model that takes a camera frame plus a text instruction and returns the next two seconds of actions. Fine-tuned on the DROID dataset, it places first on the RoboLab benchmark at 42.92% success, and ships under the FLUX Kommunity License with code and fine-tune recipes.

🔗 https://huggingface.co/blog/black-forest-labs/flux-3-action

Grok 4.7 holds its price and extends its training run

xAI shipped Grok 4.7 on September 21, its most capable model for coding and knowledge work, on a larger base model with a longer reinforcement learning run weighted toward problems that take hours. It arrives at the same price and speed as Grok 4.6, which makes the benchmark movement free for existing users, alongside xAI’s strongest safeguards to date.

🔗 https://x.ai/news/grok-4-7

Gemini 3.8 Flash TTS moves from voice presets to prompting

Google’s Gemini 3.8 Flash TTS and Flash-Lite TTS arrived September 23 as the company’s most expressive speech models yet. Rather than choosing from static presets, you can build a voice from a natural language prompt or replicate one from 30 seconds of audio, with consent checks and SynthID on the output, across the Gemini API and AI Studio.

🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/

🚀 Tools & platforms

Microsoft rebuilds Copilot around Home, Code and Autopilot

The rebuilt Copilot app on September 25 merges Chat and Cowork into Home, brings Word, Excel and PowerPoint in as Office in Copilot so drafts come out as live editable documents, adds Code so people who do not write code can build small tools from a description, and introduces a persistent Autopilot agent. Cowork, Code, Autopilot and the frontier models all move to usage based billing.

🔗 https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/

Gemini 3.8 Live pairs live dialogue with a Live Avatar

Gemini 3.8 Live with Live Avatar on September 24 gives enterprise agents a visual presence instead of a text or voice only interface, processing visual and audio input at the same time with low latency streaming video, native speech to speech across 97 languages, and background tool execution that does not interrupt the conversation. It is available in Gemini Enterprise.

🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/

Google Cloud API Gateway becomes an MCP server

In Public Preview on September 24, API Gateway can act as a remote MCP server. Instead of standing up a separate MCP server that reimplements routing, authentication and quota logic a gateway already handles, teams annotate the OpenAPI spec they already deploy and its REST operations become agent ready tools on a single endpoint.

🔗 https://developers.googleblog.com/turn-your-rest-apis-into-mcp-tools-with-google-cloud-api-gateway/

Cohere opens Compass Cloud as managed retrieval

Compass Cloud entered private beta on September 25, exposing the search and retrieval engine that has mostly powered North as a managed service through APIs, an MCP server or a Python SDK. Cohere claims nDCG moving from 64.8 to 81.1, and self hosted deployments stay available for privacy constrained buyers.

🔗 https://cohere.com/blog/compass-cloud-beta

OpenAI overhauls prompt caching for the GPT-6 family

Also September 22: higher cache hit rates by default, discounts of up to 90% on cached input tokens, and new dashboards and diagnostics for builders running persistent agents. Reused prefixes stay eligible inside a 30 minute window, which is the part that matters most for agents resending the same context on every request.

🔗 https://openai.com/index/better-prompt-caching-for-gpt-6/

💬 Community is talking about

Altman addresses the UN Security Council on AI

On September 23 Sam Altman became one of the first AI leaders to address the UN Security Council, arguing the technology should expand human agency and naming the two ways rapid progress goes badly: losing control to systems that move faster than institutions, and concentrating too much power in too few hands. He pushed for international frontier standards, incident reporting, and keeping the most capable systems under human control.

🔗 https://openai.com/index/sam-altman-un-security-council-remarks/

Claude agents report an enzyme system that looks like CRISPR

Anthropic’s new life sciences group ran roughly 950 Claude agents for 21 hours and 210 million tokens across a large DNA database and reported an independently discovered enzyme system with properties reminiscent of CRISPR, kept separate from anything previously published. The scientists involved gave high level direction only. It is the most concrete discovery claim any lab has attached to an agent swarm so far, and it will be worth watching whether it survives external replication.

🔗 https://www.anthropic.com/news/claude-discovers-novel-enzyme-system

Claude computes a nine-loop scattering amplitude for about $100

Two Anthropic physicists used Fable 5.1 inside Claude Science to compute the nine-loop MHV six-particle amplitude in planar N=4 super Yang-Mills, a calculation never done before, for roughly $100 in compute. Lance Dixon at SLAC validated the result independently and the data is published. Read it as executing a human developed recipe rather than inventing the method.

🔗 https://www.anthropic.com/research/yes-claude-can-do-nine-loops

NVIDIA, DeepMind and EMBL-EBI open 2,800+ viral protein structures

A coalition including NVIDIA, Google DeepMind, EMBL-EBI, CEPI and several universities released an open dataset of predicted 3D structures for the protein complexes of more than 2,800 viruses on September 24, inferred with AlphaFold2 optimized through NVIDIA’s BioNeMo. Roughly 30% of the interactions were new to science, and the pipeline is open too.

🔗 https://blogs.nvidia.com/blog/open-protein-dataset/

---

That’s the week. All stories above link to the original source — no aggregators in the chain. Catch the next digest next Monday, or browse the full AI tools directory while you’re here. Got feedback? Tell us.