Frontier AI, week of 23 October 2026

Two busy weeks. The big labs all shipped, agents moved from “chat window” to “works while you sleep”, and regulators started asking for documents. Here is what happened, grouped by theme, with sources.

Models and releases

Google makes Gemini 4 Argon generally available

Google DeepMind released Gemini 4 Argon on 13 October to developers, enterprises and Gemini app subscribers. Google pitches it for long, multi-step work in coding, finance, law, cybersecurity and science, and says it ranks first on the Vals Index. It launched at an introductory $2 per million input tokens and $10 per million output tokens, which Google Cloud says rises to $4 and $20 after 30 November 2026.

The truth: The internal examples (a big Chromium refactor, faster open-source chemistry code) are more telling than the leaderboard claim, because they are checkable. Note the price doubles in December: budget for the real number, not the launch one.

Source: Google: Gemini 4 Argon is now available for everyone · Google Cloud pricing note

Meta launches Muse Wonder and Muse Wonder Full

On 12 October Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models, aimed at agent work such as insurance claims, leases and app building. Meta highlights “Active Memory” across conversations and “Active Focus” for sticking with hard tasks. Muse Wonder is available through the Meta Model API and Muse Code; a full safety report is promised “soon”.

The truth: Meta says it rebuilt its AI effort over fourteen months, and this is the payoff it wants you to see. Until the promised safety report and independent tests land, “most capable to date” is a claim about Meta’s own line-up, not the field.

Source: Meta: Introducing Muse Wonder and Muse Wonder Full

Anthropic ships Claude Haiku 5.5 and cuts prices

Just before our window, on 7 October, Anthropic released Claude Haiku 5.5, its small, fast model for high-volume jobs like summaries, classification and acting as a sub-agent for bigger models. Anthropic says it costs around 75% less to run than Haiku 4.5 on average, and it also halved Sonnet 5.5’s cache-read price.

The truth: The headline is price, not intelligence. Most real AI spend goes on lots of small, boring calls, and that is exactly where this competes.

Source: Anthropic: Introducing Claude Haiku 5.5

Research

Does “latent chain-of-thought” actually reason?

A paper posted to arXiv on 15 October tests models that reason in hidden states instead of written steps, across 21 model-task setups. On logic tasks, the COCONUT method often already “knows” the answer before its latent chain starts, and deleting the chain barely changes predictions; on arithmetic, all tested recipes really did depend on the chain. The authors propose a taxonomy separating what a chain holds from how the model uses it.

The truth: This is the healthy kind of research: it checks whether a popular idea does what people say. Being able to read an answer out of a hidden state is not proof the model reasoned its way there.

Source: arXiv 2610.21182: Does Latent Chain-of-Thought Actually Reason?

Agents and tools

Anthropic previews Claude Duet, a proactive agent app

On 21 October Anthropic opened a waitlisted research preview of Claude Duet, a separate app (iOS, Android, web) for US Pro and Max subscribers. Built around Claude Opus 5.5, it sends a morning brief, watches connected apps every few hours, and runs tasks in its own cloud virtual machine. It must ask for approval before sending messages, paying, signing or deleting, and some actions, such as trading securities, are off-limits.

The truth: The real story is the shift from “answer when asked” to “interrupt you when it thinks it matters”. Anthropic itself admits it can misjudge urgency; the approval gates are the feature to watch.

Source: Anthropic: Introducing Claude Duet

OpenAI makes its Decisions API generally available

On 22 October OpenAI made the Decisions API generally available, built on GPT-6 Luna, for fast questions with a fixed set of answers: classify this, route that, pick the next action. OpenAI says it is up to 10x faster than asking GPT-6 Luna through the Responses API, and each request can now hold up to 400 questions, double the beta limit.

The truth: Unglamorous and important. Agents make thousands of tiny decisions; making those cheap and fast matters more to real products than another benchmark point.

Source: OpenAI: Build more responsive apps with the Decisions API

Google Cloud launches the Gemini agent for work

At Gemini at Work on 8 October, Google Cloud announced the Gemini agent: one prompt box that plans tasks, uses tools and connects to company systems, working inside Gmail, Drive, Docs and other Workspace apps. Google stresses governance features such as identity and permission controls, sandboxing and real-time spend caps.

The truth: Every lab now sells “one agent for all your work”. The differentiator is no longer the demo but the admin controls, and Google is leading with exactly that.

Source: Google Cloud: Gemini at Work 2026

Policy and safety

The FTC subpoenas OpenAI and Anthropic

The US Federal Trade Commission confirmed it has subpoenaed Anthropic and OpenAI over possible consumer-protection violations, CBS News reported on 15 October. Citing Axios, CBS says the agency wants documents on safety incidents, privacy policies and safeguards, following incidents in which AI models hacked into websites, and plans to subpoena third parties later.

The truth: This came while the White House resists new AI rules, so it is consumer-protection law doing the work that AI-specific law doesn’t. Documents under subpoena tend to say more than blog posts.

Source: CBS News: FTC issues subpoenas to Anthropic and OpenAI over safety incidents

Industry and compute

Anthropic in talks for another ~$45bn of Nscale compute

Reuters reported on 19 October, citing a person familiar with the matter, that Nscale is in talks to sell Anthropic computing power worth about $45 billion from a second building at its Monarch campus in West Virginia, using Nvidia Vera Rubin chips. Anthropic struck a similar $45 billion deal in August. The same report notes investor worries about debt-funded AI infrastructure.

The truth: These are talks, from one anonymous source, and “details could change”. The signal is scale: the frontier is now a power-and-financing business as much as a research one.

Source: Reuters via Investing.com: Nscale, Anthropic in talks for new $45 billion AI cloud deal

Google Cloud brings Nvidia Vera Rubin online

Nvidia said on 21 October that Google Cloud’s A5X instances, built on Vera Rubin NVL72 systems, reach general availability next month, with Europe to follow. Google plans to scale to as many as 960,000 Rubin GPUs across sites; David Silver’s Ineffable Intelligence was the first customer, using it for large-scale reinforcement learning.

The truth: “960,000” is a plan, not an installed number. The more interesting detail is who went first: a reinforcement-learning lab betting on learning from experience rather than more internet text.

Source: NVIDIA: Google Cloud and NVIDIA announce general availability of Vera Rubin infrastructure

OpenAI models are coming to Google Cloud

On 21 October OpenAI and Google Cloud announced OpenAI models and Codex will be sold through Google Cloud’s Model Garden, with general availability “in the coming weeks”. OpenAI says its models will also come to Google Antigravity and the Gemini agent, and that Microsoft remains its primary cloud partner.

The truth: Rivals selling each other’s models tells you where the money is: in the cloud contracts, not in exclusive model loyalty.

Source: OpenAI: OpenAI models are coming to Google Cloud

Open source and open weights

Meta releases Muse Spark Open weights

On 15 October Meta released weights for Muse Spark Open, a multimodal model with 816 billion total parameters (42 billion active) and a 1-million-token context, which it calls the most capable US open-weight model. Meta says it can run on two Apple M5 Ultra Mac Studios. The licence is the Muse Model License with “limited commercial terms”.

The truth: “Open weight” is not “open source”: read the licence before you build a business on it. Still, a frontier-class model you can run at home is a real shift in who gets access.

Source: Meta AI Research: Introducing Muse Spark Open

DeepSeek swaps in V4.1 Flash 1017

DeepSeek’s API docs say the deepseek-flash model name is now served by DeepSeek-V4.1-Flash-1017, with the calling method unchanged. Older Flash names still work but are routed to the new model and billed at the Flash price.

The truth: A quiet in-place upgrade. Good for users, but if you depend on stable behaviour, a model changing under the same name is something to test, not assume.

Source: DeepSeek API docs

Sources were read on 23 October 2026. Spotted an error? That’s the point: tell me.