Meta ships its first frontier model; Microsoft bets on Maia 300

Meta Superintelligence Labs released Muse Wonder and a larger Muse Wonder Full on Monday, the company's first models aimed at the top of the field rather than at speed. Meta's own testing puts them level with Claude Fable 5.1 and GPT-6 Astra, and ahead of both on a memory benchmark, a skill Meta calls Active Memory: carrying what matters about a user across conversations for weeks. Muse Wonder costs $4 per million input tokens and $20 per million output through the Meta Model API, close to Opus 5.5 pricing, while Muse Wonder Full is limited to Muse Code subscribers and is not offered through the API. Both were trained at Meta's Prometheus data center in Ohio and will later power the Muse agent, starting with subscribers. Meta has not announced open weights for either model.

Microsoft announced Maia 300, the third generation of its in-house AI chip, and said it will scale the Maia family to deployments of several gigawatts starting in 2027, making it the main system that serves AI answers across Microsoft products. The company expects Maia 300 to produce 40 percent more tokens per megawatt than leading GPU systems, and up to seven times the rack throughput of Maia 200 in a 96-chip configuration. It is built from chiplets on TSMC's N2P process, with eight HBM4 memory stacks. Maia 200 already serves Copilot and runs OpenAI models, with GitHub Copilot and Microsoft Foundry to follow by year end. The Verge framed the move as Microsoft becoming less dependent on Nvidia.

Epic Games, the maker of Fortnite and the Unreal Engine, is moving its Disney project to an "AI first" model in which models such as Anthropic's Claude write all of the code and generate the artwork, while developers manage and guide the output, Bloomberg reported from internal messages. The Marvel and Star Wars shooter, part of a $1.5 billion partnership, had been due out in March 2027 and is now delayed indefinitely. Chief executive Tim Sweeney confirmed "a huge pivot" on social media and said the aim is to build more, not to do the same work with fewer people. Epic laid off more than 1,000 employees in March.

Microsoft president Brad Smith told a BBC podcast that the moment has arrived for global AI safety rules, noting that milk and washing machines face more oversight than frontier labs. He named Britain as best placed to lead, because its AI Security Institute already tests major models before release and Washington does not view London the way it views Brussels. Smith called the voluntary pledge labs signed at the White House last month a good step, but said only responsible companies follow voluntary rules. He backed a legal duty to report breaches by rogue AI agents and approval by outside evaluators before release. The institute confirmed on Monday that it can still test the latest models, after reports that the White House asked labs to hold them back.

Berkeley computer scientist Ben Recht took apart one result from the 722 math papers OpenAI published last week, a claimed algorithm that beats the fast Fourier transform, one of the most widely used routines in signal processing. In his argmin newsletter he argues the proof assumes exact arithmetic on complex numbers whose size grows faster than any polynomial, so storing them would take memory that grows exponentially with the input. Lower bounds dating to 1973 already show that realistic limits on those numbers rule out such a speedup, so he calls the result "less a breakthrough and more a malicious exploit" of the model of computation. Recht did not check whether the proof itself holds. OpenAI had already withdrawn three papers from the set after finding a sign error.

The benchmarking company Vals ran GPT-6 Sol and Claude Opus 5.5 on 50 full web-app builds, once as single agents and once as teams where a lead agent hands work to up to five subagents. Teams cost 1.8 to 5.1 times as much, and only one of four comparisons showed a real gain: Sol at medium effort, up 7.3 points. For Sol, raising its reasoning effort helped more than adding a team. The most expensive setup, Opus at maximum effort as a team, scored best at 93.2 percent but cost about $122 and nearly three hours per app. Most of the extra spend came from each subagent re-reading its own copy of the specification on every call.

A paper accepted at NeurIPS 2026, led by Andis Draguns, shows a model can carry an encrypted network inside its own weights that unlocks only when it sees a secret trigger. Under standard cryptographic assumptions the hidden circuit is provably hard to find or draw out, even for someone holding the full weights. The team tested it on a chemical reaction task standing in for weapons-relevant skills, and fine-tuning, adding noise to the weights, searching for the trigger and adversarial training all failed to surface it. The authors say this sets a limit on safety audits that rely on drawing out a capability, and they released their models so others can work on defenses.