The Signal — October 13, 2026

Meta released what it calls its most capable models yet, with features intended to keep them working through longer tasks. Britain says its AI Security Institute has kept access to test the latest models before release, and Microsoft plans to make its own chips the main system for running AI across its products and services from 2027.

Meta releases Muse Wonder and Muse Wonder Full for agent work

Meta introduced Muse Wonder and Muse Wonder Full on October 12, calling them its most capable models to date. Developers can use Muse Wonder through the Meta Model API and Muse Code, Meta's coding product, while Wonder Full is available only to Muse Code subscribers. Meta says Wonder will soon power its Muse consumer app, starting with subscribers. The company describes two features aimed at longer tasks: Active Memory carries relevant details across conversations, while Active Focus keeps the model working toward a goal when a task gets hard and prompts it to check in before taking important actions. Meta says a full report with additional safety evaluations "will be available soon."

Artificial Analysis, which runs an independent model leaderboard, scored Wonder 52 and Wonder Full 56 on its Intelligence Index at maximum reasoning effort, up from 48 for Muse Spark 1.3, which came out six weeks earlier. Wonder Full placed in the top three on six of the ten tests in the index. It took first place on CritPt and GDP.pdf (33% each), but the leads are narrow, with the nearest rivals at 31 to 32%. On GDPval-AA, its result falls within the confidence intervals of Claude Opus 5.5 and Sonnet 5.5. Its biggest gain over Spark was on Terminal-Bench 4.0, from 33% to 57%.

The models cost more to run. Wonder is priced at $4 per million input tokens and $20 per million output tokens, against $1.25 and $4.25 for Spark 1.3. Artificial Analysis measured $4.52 per index task compared with $1.60 for Spark, mostly because Wonder reads a large amount of input. General knowledge remains a relative weakness, though Artificial Analysis found Wonder Full among the least likely of the models it tested to make up an answer. "State of the art" is Meta's description, Active Memory and Active Focus are Meta's claims about how the models behave, and the scores come from one benchmarker. Our October 10 edition covered a New York Times report on the launch of Meta's Muse agent; Meta says Wonder will soon power that app.

Sources: Meta · Artificial Analysis

Britain says its AI Security Institute still gets early access after a reported White House reversal

In a speech at Imperial College London on October 12, UK AI minister Kanishka Narayan said the government had spent recent weeks working with the US government and AI developers "to retain access for Britain's AI Security Institute to the most powerful AI models before they are released." He confirmed that the institute's scientists "are now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon." The institute gets this early access under voluntary agreements with the companies.

The day before, Reuters relayed a Telegraph report that the Trump administration had reversed an earlier request that developers such as OpenAI and Anthropic not share new models with the institute before release. Spokespeople for Trump, OpenAI and Anthropic did not immediately respond to Reuters. In September we held Politico's original report of that request because it rested on anonymous sources; the minister's speech is the first on-record account of where access stands.

The speech leaves gaps. Reuters noted that Narayan did not list all the models the institute is evaluating and that ministers have never confirmed a wider loss of access. It also noted a September letter in which the institute said Anthropic had made clear that no organization outside the US had access to its Mythos 5.1 model. Narayan named only a Google model. "When you're reliant on someone else's technology, you have to be clear-eyed about the permanence of any outcome," he told reporters, while saying he was confident Britain would be able to test the latest models. The reversal is the Telegraph's account, and we did not find a public US directive confirming it.

Sources: GOV.UK · Reuters · Reuters via U.S. News

Microsoft unveils Maia 300 and plans to make its own chips its main inference system from 2027

Microsoft announced Maia 300, the third generation of its in-house AI chip, on October 12. In a post by Azure chief Scott Guthrie, the company said it will "scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services." Inference is the work of running trained models for users, as opposed to training them. Maia 300 begins production deployments in 2027.

Microsoft expects Maia 300 to produce 40% more tokens per megawatt than leading GPU-based systems, and a rack of 96 chips to deliver up to seven times the token throughput of a Maia 200 rack. According to Microsoft, the compute chiplets are made on TSMC's N2P process. Each chip has eight HBM4 memory stacks with 18 TB/s of peak bandwidth within a 950-watt power budget, and a full rack draws 167 kilowatts. The previous generation, Maia 200, announced in January, is already serving Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. That chip runs OpenAI models, Microsoft's own MAI models and open-source models.

Satya Nadella called the Maia expansion "our biggest bet yet on our own silicon," GeekWire reported, and Microsoft AI chief Mustafa Suleyman told The Verge that running Microsoft's models on Maia is "a critical focus." The efficiency figures are Microsoft's projections against GPU systems it did not name, and no outside group has measured them. Multi-gigawatt scale is a plan for next year onward, so how far it reduces Microsoft's purchases of Nvidia and AMD chips is not yet known.

Sources: Microsoft · The Verge · GeekWire

On the Editor's Desk

A Microsoft Research paper found that long-running research agents do better when they periodically restart from a fresh conversation while keeping their files. We left it out because we have covered closely related agent research several times in the past week. We also held Nvidia's contribution of Vera Rubin rack designs to the Open Compute Project because the efficiency benefits are Nvidia's own claims. Italy's new AI Act enforcement decree remains on hold while we confirm when it takes effect.