September in brief

Anthropic shipped three new Claude models and OpenAI shipped four from the GPT-6 family. Google announced Gemini 4 Argon on September 30, the last day of the quarter, just in time to count it as a Q3 launch. And Meta and OpenAI both launched personal agents, joining Grok Bot and Gemini Spark.

- The personal agent field got crowded. Meta launched Muse, OpenAI started rolling out dots to paid ChatGPT plans, and xAI took Grok Bot, out since August, into Tesla cars.

- Prices fell, and not only at the top. Anthropic cut cache read prices 75 percent and priced Opus 5.5 20 percent below Opus 5. GPT-6 Sol and Luna came in at about half their predecessors' prices, and xAI kept Grok 4.7 at $2 and $6 per million tokens.

- Safety findings changed what shipped. OpenAI paused tool use after an agent escaped its sandbox, then dropped the GPT-6.1 Astra update. Leaders of six AI companies signed a voluntary White House pledge calling for outside auditors at frontier labs.

- Models did serious math. Claude turned Wiles's proof of Fermat's Last Theorem into machine-checked Lean, OpenAI said an unreleased model solved the Navier-Stokes Millennium Prize Problem (a credit dispute followed), and agents wrote the GPU code behind two RSA factoring results.

What launched

33 products and models, earliest first.

- Same underlying model, two safety configurations: Fable 5.1 ships generally available, Mythos 5.1 only through trusted access programs for cybersecurity and life-sciences work. - Anthropic · Model · Source

- Its third Flash-tier release in six weeks, plus 3.8 Flash Cyber, gated to vetted defenders through Google's new Fairwind Program. - Google DeepMind · Model · Source

- Two days after saying Astra had crossed the Critical cybersecurity threshold of its own Preparedness Framework, OpenAI shipped it to a first set of organizations, with Plus, Pro, Business and Enterprise, the API, Azure and Bedrock to follow. - OpenAI · Model · Source

- Muse books appointments, shops, and works through multi-step tasks from a dedicated cloud VM with its own browser, built under chief AI officer Alexandr Wang. - Meta · Product · Source

- An update to OpenAI's image model, rolling out across ChatGPT, ChatGPT Work and Codex with two new API variants (GPT-Image-2.5 Flare and Sunburst), cutting generation latency up to 50 percent versus Images 2.0 while improving multi-turn edit reliability. - OpenAI · Product · Source: 9to5Mac

- DeepSeek released V4.1 Flash under an MIT license, a 552-billion-parameter multimodal mixture-of-experts model with a new causal encoder-decoder architecture that cuts its KV cache footprint to a quarter of the prior generation's. - DeepSeek · Model · Source

- The voice model that listens and speaks at once, shipped in ChatGPT in July, became available to developers at $0.05 a minute. - OpenAI · Model · Source

- Cognition says its new in-house model for Devin scores within a point of Claude Fable 5.1 on FrontierCode, a benchmark Cognition maintains itself, at 64 percent lower cost. - Cognition · Model · Source

- A Project is a cloud-hosted coordinator that keeps context for months, hands coding to subagents, and wakes up on a schedule, a Slack channel or a new pull request. - Anysphere · Product · Source: Cursor

- A version of ChatGPT Work on GPT-6 Astra for investment banking and equity research, with data from Daloopa, PitchBook, LSEG News and Crunchbase hosted by OpenAI and firm templates for models and pitchbooks. - OpenAI · Product · Source

- The kimi-for-coding ID now serves K2.8 Preview, which Moonshot says performs close to K3 while thinking more efficiently, with up to 1M tokens of context on every plan that includes it. - Moonshot AI · Model · Source

- Both route each task through a pool of models that Sakana only partly discloses. - Sakana AI · Model · Source

- iOS 27 and Apple's other 2027 software releases brought the rebuilt Siri to supported devices in English, with daily limits on server-side features and paid expanded access promised later. - Apple, Google · Product · Source

- Two speech-to-speech models: a cheaper Live built for scale, and an Extended Thinking version that reasons while it talks and runs tool calls in the background. - Google · Model · Source

- Anthropic began rolling out, over several weeks, a Claude for Pro and Max subscribers with no separate chat and agent modes: anything Cowork could do became available from any conversation, with tasks running in the cloud. - Anthropic · Product · Source

- GPT-6 Astra with a legal search index of more than 230 million URLs, including CourtListener's case law, and instructions for legal analysis and writing. - OpenAI · Product · Source

- One model for text-to-image and editing, with a 7 billion parameter image generator, native transparent layers and up to 10 reference images. - Alibaba · Model · Source: Qwen

- A new, larger base model than Grok 4.6's, trained longer on tasks that take hours, at the same $2 and $6 per million tokens. - xAI, SpaceX · Model · Source: SpaceXAI

- The first of the Claude 5.5 models, priced at $4 and $20 per million tokens, 20 percent below Opus 5, and pitched as performing at Fable 5.1's level on most work. - Anthropic · Model · Source

- The cheaper tiers of the GPT-6 family, trained with methods similar to Astra's, arrived in ChatGPT Work, Codex and the API at about half their predecessors' promotional prices: $2 and $10 per million tokens for Sol. - OpenAI · Model · Source

- Two omnimodal models, a 1.02 trillion parameter Pro with 42 billion active and a smaller Flash, released with Xiaomi's reinforcement learning code. - Xiaomi · Model · Source

- Drivers can hand Grok multi-step errands by voice, such as a coffee order or a reservation, and use Connectors to work through email and calendars. - Tesla, xAI · Product · Source

- Fireworks put Ember-1, a post-trained Kimi K3 variant, on its serverless API for an initial two-week research preview. - Fireworks AI · Model · Source

- Google made Live Avatar available in Gemini Enterprise, pairing Gemini 3.8 Live speech with generated video personas that process audio and visual input while calling tools. - Google · Model · Source

- An early preview lets Gemini phone US businesses from the user's own number to book, check stock or ask for quotes, opening every call by saying it is an AI on a recorded line. - Google · Product · Source

- Anthropic released its second Claude 5.5 model for everyday coding and document work. - Anthropic · Model · Source

- Eleven v4 added finer voice direction, multi-speaker context and support for more than 90 languages; v4 Turbo was tuned for conversational agents. - ElevenLabs · Model · Source

- Meta added business skills and connectors to its existing Muse agent, including access to Facebook and Instagram business accounts and tools such as QuickBooks, Shopify and Stripe. - Meta · Product · Source

- Each dot runs on its own cloud computer and can work across connected apps under configurable action rules. - OpenAI · Product · Source

- An upgrade to GPT-6 Sol that OpenAI says approaches Astra on some coding and computer-use tests at lower cost. - OpenAI · Model · Source

- A shared home in ChatGPT for pages and files, where teammates, ChatGPT and a user's dot edit the same documents. - OpenAI · Product · Source

- Participating tools, 16 at launch including Devin, Notion and Vercel, can now draw AI requests from a Plus or Pro subscriber's ChatGPT Work and Codex allowance instead of an API key, under per-app weekly caps. - OpenAI · Product · Source

- Google's new frontier model, with a 1 million token output limit, went first to partners in its Fairwind cyber-defense program while Google took part in the US government's voluntary pre-release access process. - Google DeepMind · Model · Source