DeAI Daily Brief — 13 October 2026

Today in DeAI: Meta ships frontier-class models with no public weights, Mistral's open-weights week arrives, Microsoft shows clearing agent context buys 66% better scores, the White House restores UK safety-tester access, and an AI math system passes peer review in Science.

Meta Muse Wonder arrives closed-weight, and the per-task bill nearly triples

Meta Superintelligence Labs launched Muse Wonder and Muse Wonder Full on October 12, which Meta describes as its most capable models yet. They add Active Memory and Active Focus, 1M-token context, and multimodal input. Meta reports 56 on the Artificial Analysis Intelligence Index for the Full tier, up from 48 for Muse Spark 1.3, with the biggest gains on agentic coding. Nothing is downloadable, and per Artificial Analysis, Muse Wonder costs $4.52 per task versus $1.60 for its predecessor, at $4/$20 per million input/output tokens. Why it matters: the frontier tier is converging on agentic workloads priced per task, while the open-weights door that Meta's earlier releases left open stays shut here. Full story: our lead coverage (Meta, The Decoder).

Mistral Large 4 weights week: two dates, one confirmed

Mistral's own words promise the ~1T-parameter, 52B-active Large 4 weights "by the end of the month"; the Hugging Face upcoming-release page shows an Oct 31 ETA, while the Oct 27 date circulating in secondary coverage remains unconfirmed and the license is still unnamed. Why it matters: if the license lands Apache 2.0 like Mistral Large 3, a frontier-class 1T mixture-of-experts model becomes self-hostable and servable by any marketplace on day one; if not, the preview pricing is the whole story. Background: our Oct 7 coverage.

Microsoft's Fork-and-Flush: clear the chat, keep the workspace, gain 66%

Microsoft Research found autonomous research agents get stuck in "idea basins" as conversation context accumulates, plateauing at widely different scores on identical compute. Their Fork-and-Flush method (branch into parallel runs, wipe each branch's chat, keep the workspace) reached a 0.78 normalized average score versus 0.47 for single runs, a 66% relative gain across 13 long-horizon tasks (The Decoder's writeup). Why it matters: context management is moving from implementation detail to a first-order determiner of agent quality and token spend, on any inference tier. Practical grounding: our agent-API guide.

White House restores UK AISI's frontier-model testing access

Per The Telegraph, the White House reversed its September directive and restored the UK AI Security Institute's pre-release access to models from Anthropic, Google DeepMind, and OpenAI as of October 11, ending a 17-day freeze after UK ministerial lobbying. Why it matters: AISI is one of the few external evaluation paths frontier models get, and its interruption shows third-party evaluation access is a policy variable, not a guarantee. Related: our White House and Anthropic coverage.

AlphaProof Nexus solved nine open Erdős problems, peer-reviewed in Science

Google DeepMind's AlphaProof Nexus autonomously solved nine of 353 open Erdős problems, two open for 56 years, plus 44 OEIS conjectures, at a few hundred dollars of inference per problem, and the work passed peer review in Science (The Decoder, the arXiv report). Why it matters: the same capability claims ran through two opposite verification disciplines this week; the community's trust went to the one that submitted to review, which is the same verification standard provider claims are held to here.

Watching tomorrow

StepFun's full Step 5 weights are claimed for Oct 15 on Hugging Face, an unverified relay so far; verify the HF org page before treating the drop as fact. The Mistral Large 4 weights window stays open through Oct 31.