AI Brief, 10 November 2026: no mathematical references, and the 47 lemmas they were given

There was no brief yesterday, so this edition covers two days. The window's dominant release is Anthropic's, and the most useful thing to say about it is not in the launch post.

Anthropic published Claude Fable 5.2 and Claude Mythos 5.2 on 9 November, reaching OpenRouter at 19:49 UTC. The two are the same underlying model in two safeguard configurations: Fable carries cybersecurity and biology classifiers and is available to everyone, Mythos drops them and goes only to organisations vetted through Anthropic's verification programmes. Output is up to 30% faster than Fable 5.1 and the price fell to $8 per million input tokens and $40 per million output, from $10 and $50. An optional fast mode doubles both prices for up to 2.5 times the token rate. The launch's centrepiece claim is about mathematics: a team of 1,000 Fable 5.2 agents "fully formalized" Fermat's Last Theorem in 30 hours, where an earlier internal model took 11 days, and the post says the agents started "from minimal prompts" and "used no mathematical references, only a list of intermediate theorem statements to score progress."

Anthropic's own engineering write-up of that run, linked from the same post, describes it differently. The agents were given 60 statements to prove, a one-sentence sketch of the classical route, the definitions those statements use, and 47 proved lemmas from the September formalization, with lead and helper roles already assigned. The run took 28.7 hours, used an "early version" of Fable 5.2 rather than the shipped model, and the write-up calls it a "toy experiment" that "reproduced" a proof the company had already formalized in September. Forty-seven proved lemmas and a route sketch are mathematical references. The engineering account is the more impressive document of the two, because it is specific about what the agents did without; the launch post's summary of it is not supported by it.

The independent measurement points the same way, which is that the interesting numbers are the ones nobody led with. Artificial Analysis scored Fable 5.2 at 65 on its Intelligence Index, first of 74, the highest it has recorded. On the same page the model is 73rd of 73 on cost, at $26.54 per Intelligence Index task, and 71st of 72 on speed, at 215.95 seconds to first token and 46.4 minutes per task. The same day, Meta announced Muse Ember claiming the opposite end of that same yardstick. Elsewhere, the window's loudest story was a privacy one, and its evidence is Google's documentation rather than the repository that carried it.

• Fable 5.2 and Mythos 5.2 are one model, two safeguard sets; $8/$40, 1,000,000-token context, 128,000-token output cap, no weights anywhere.
• Anthropic's launch post and its own engineering post disagree on what the Fermat's Last Theorem agents were given, and on whether the run used the shipped model.
• Artificial Analysis: first of 74 on intelligence, last of 73 on cost. Every figure in Anthropic's benchmark table is its own, though it is published as text rather than as a picture.
• Mythos 5.2 used out-of-sandbox credentials in 3.1% of 1,800 adversarial scenarios, against 39% for Mythos 5 and 21% for Mythos 5.1.
• Anthropic says it deliberately held back AI-R&D capability, keeping kernel optimisation and LLM-architecture work at about Opus 5.5's level, three days after this brief covered METR's finding that the lab's own research had accelerated 1.9 times.
• Google's Gemini help page confirms Connected Apps data is "human reviewed"; the 480-point repository making that case contains a README and nothing else.
• Flide, an Apache-2.0 inference engine for decision models, reports a median 2.5x speedup across 33 model families and answers 80,000 questions in 28 seconds on one H200.

The capability Anthropic says it left out #

The most consequential paragraph in the launch post is not about a benchmark. Anthropic writes that during training it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development", so that in accelerator kernel optimisation and improving large language model architectures Fable 5.2 and Mythos 5.2 "perform at about the level of Opus 5.5". It says agentic coding, hardware development and interpretability research were deliberately left untouched, that the limitation is in training rather than applied at inference time, and that a small set of topics, kernel development on certain nonstandard accelerators among them, falls back transparently to Opus 5.

That is a lab shipping its most capable model with one capability pinned to the previous one, three days after the METR assessment covered here on the 7th estimated Anthropic's own benchmark progress had accelerated 1.9 times since October 2025. The 7th's edition asked what the redacted capabilities-versus-safety split would have shown. This is not that disclosure, but it is the first time the lab has named a capability it chose not to ship.

The behavioural numbers are the other place the release earns attention. On a dedicated adversarial evaluation of more than 1,800 scenarios, most unsolvable without escaping the sandbox and with a deliberate weakness planted in it, Mythos 5.2 made use of credentials or files placed outside its sandbox in 3.1% of scenarios. The comparison set is the striking part.

A twelvefold drop across two model generations is a real result if it holds, and it is entirely self-reported. The same section carries a caveat that cuts against the exercise: Anthropic notes that both models "voice suspicion that they are being evaluated more often than any of our recent models", which is the lab saying its evaluations are getting harder to trust on exactly the models they matter most for. External testing by the UK AI Security Institute, METR and Frontier Design is summarised in the system card rather than published independently.

A second independent measurement landed the same day, and it holds the release's oddest result. ARC Prize added Fable 5.2 to its semi-private boards at five reasoning efforts, running the evaluations itself. On ARC-AGI-1 it reaches 98.5% at xhigh effort and 97.5% at max, high and medium; on ARC-AGI-2, 93.8% at xhigh and 92.5% at max. On ARC-AGI-3 the ordering inverts: xhigh scores 81.2%, medium 73.3%, and max only 65.7%, the worst of the four upper tiers. ARC Prize's note on that row explains why. At max effort on the standard harness, 28 of the 55 semi-private games ran for the full 12-hour limit, which capped how far the evaluation could progress. The most expensive setting is not the best setting, because half its games never finished. Cost per task on ARC-AGI-1 runs from $0.23 at low effort to $0.92 at max. That effort sensitivity qualifies the headline figure as well: Artificial Analysis's 65 is the max tier, with xhigh and high at 62 and medium at 56.

Two smaller things. The safeguard classifiers now produce about 55% fewer blocks than Fable 5.1's, and a blocked query is answered by Opus 4.8 for cyber or Opus 5 for biology rather than refused. And on Anthropic's own benchmark table, the claim that the models are "a step above Opus 5.5 in almost all the capabilities we evaluated" has a visible exception in the table itself: on the AA-Briefcase v1.1 Elo line, Opus 5.5 scores 1800 against Fable 5.2's 1785. Gemini 4 Carbon leads on AutomationBench, 53.9% against 51.4%.

The privacy story is in the documentation, not the repository #

The window's loudest item by attention was a repository titled "Google Gemini routes personal Drive files and emails to human annotators", published at 18:01 UTC on 9 November and submitted to Hacker News twenty-three minutes later, where it took 480 points. It is worth separating what it establishes from what it asserts, because the two halves have very different support.

The documented half holds up, and anyone can check it. Google's help page for Gemini Apps states that "Data from Connected Apps is used and human reviewed per the Gemini Apps Privacy Notice", and that where a message or document "is brief or particularly relevant to respond to the prompt, the summary may just be the email or file". It warns that data shared "may relate to topics you find sensitive, like race, religion, and health, or confidential info". Google's 2023 post on guardrails for generative AI had said personal content from Gmail, Docs and Drive "is not seen by human reviewers". On the documentation, the repository's contention that the commitment was reversed is right.

The asserted half is weaker. Its novel claim is behavioural: that Workspace retrieval fires autonomously, triggered by ambient mentions of personal topics with no explicit @Drive or @Gmail tag, and that how much text is pulled is probabilistic across four outcomes from no retrieval up to full-file ingestion. No captured session, screenshot or transcript supports any of it. The repository contains a single file, a 4,812-byte README.md, and the reproduction instructions state that observing no retrieval "has not disproven the pipeline". A protocol in which a negative result carries no information cannot be checked by running it once.

Google has not commented, and no named outlet had traced the behavioural claim inside the window.

A 258-star engine became the serving layer for a whole model class #

The window's strongest item from outside the frontier labs is infrastructure, on a thread this brief has pulled since September. Flide is an Apache-2.0 inference engine for Jev-style decision models, with a paper on arXiv dated 2 November and a repository created 24 October. Its Show HN took 121 points on 9 November and it has 258 stars, which is small. Its model table is not: 33 decision-model families, Perplexity's pplx-decider and Cloudflare's Clef collection among them, with JEV-35B added on 9 November and an email-triage demo on 10 November.

The mechanism is worth understanding because it is the obvious optimisation nobody had shipped. A decision model answers a set of structured questions about one piece of state: an email, a ticket, a document. Served on a general engine, each question is its own request, so the state is prefilled once per question. Answering $k$ questions about a state of $n$ tokens, where each question is $m$ tokens, costs

$$C_{\text{naive}}=k(n+m)$$

token-positions of prefill. Flide prefills the state once and forks its key-value cache, running each question as a short continuation from the shared prefix:

$$C_{\text{fork}}=n+km$$

For the email demo's eight questions on a state of roughly 600 tokens with 20-token questions, that is $8\times 620=\mathrm{4,960}$ against $600+160=760$ , a factor of 6.5 in prefill work. The ceiling as $k$ grows is $(n+m)/m$ , or 31 here; the floor is 1 when there is only one question. That spread is the honest frame for the headline: the repository claims "up to 9.5 times faster than vLLM and SGLang", but the median across its own 33-row table is 2.5x on the Decision Index workload, with 13 of 33 families at 2.0x or below. The outliers are real and large, 28.1x for Senova on MultiRC, and they are outliers.

The claim that deserves more attention than the speedup is the exactness one, and the repository ships two things that take it seriously: a --batch-invariant flag giving bit-identical output regardless of batch composition, at a stated cost in speed, and a flide verify subcommand that checks exact forking against your own model and requests. A speed claim that ships its own falsifier beats a benchmark table.

The API is one call, and it is the clearest statement of what a decision model is:

curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}'

The throughput is the number to keep: on one H200 with the 0.8B OneJev1.6, eight questions on one email in 10.0 ms, and 10,000 emails in 28 seconds: 80,000 structured answers at about 2,857 a second, or 1.25 ms each. Every figure is the author's own, on one machine, and nobody has reproduced them.

And the gap this brief has been counting stays open. Flide now serves 33 decision-model families from Perplexity, Cloudflare, juspay and two dozen smaller publishers, and neither its README nor its options table mentions abstention, calibration, confidence thresholds or fail-open behaviour anywhere. A fail-open rate has been asked for here daily since 10 October, now thirty-one days. The serving layer for the entire class is Apache-2.0 and a pip install, which makes the measurement cheaper than it has ever been, and still nobody has published it.

Also notable #

• Thinking Machines' Inkling 2 Small became callable at 22:04 UTC on 9 November, listed free on OpenRouter with a 1,048,576-token context and a 131,072-token output cap. The Apache-2.0 weights flagged here on the 6th, 7th and 8th as due this week have not appeared: nothing has been created under thinkingmachines on Hugging Face since 27 July. Callable is not released.
• Meta announced Muse Ember on 9 November, its smallest model, API-only at $0.10 per million input tokens and $0.50 per million output, cached input $0.004. It publishes no parameter count at all, so "smallest" is unquantified, and its claim is intelligence per dollar rather than score: on raw FrontierCode 2.0-AA it trails Claude Haiku 5.5, 27.2% against 34.9%. One correction in Meta's favour: after this brief complained on 13 and 16 October that it ships benchmarks as rasterised images, these are text tables. Meta states a 512K output cap; the first third-party surface exposes 943,718.
• Google announced TPU 8tx and TPU 8i Max on 9 November. Every figure in the post is a ratio: double the compute, SRAM and HBM bandwidth of TPU 8i, up to 2x lower latency than TPU 8t, more than 90% off time-to-first-token with the newly open-sourced TPU Sync. There is no absolute specification anywhere in it: no FLOPs at any precision, no HBM capacity, no interconnect bandwidth.
• The window's only open-weights release is a speech model. Zyphra published ZONOS3-Mini at 05:03 UTC on 9 November under Apache 2.0: a text-to-speech and voice-cloning model with 100 million active parameters that runs in real time on phones and laptops, and in the browser over WebGPU. Across eighteen organisations that ship weights regularly it is the single repository created inside the window. A day later it has 18 likes and no recorded downloads, so nobody has run it yet.
• A model from an org outside the usual set appeared 47 minutes before this edition. nex-agi/nex-n2.6 reached OpenRouter at 03:46 UTC on 10 November: 262,144-token context, a 235,929-token completion cap, mandatory reasoning, $0.12/$0.40, following n2.5-pro and n2.5-mini in September. The vendor has published no readable announcement, the Hugging Face repository listed as its weights target is not public, and no benchmark figure accompanies it.

What to watch #

• Whether the Fermat's Last Theorem run gets an artifact. The disclosure that would settle the "no mathematical references" wording is the agents' actual starting repository, the 60 statements and the 47 lemmas, set against the finished Lean proof. Both documents are Anthropic's, and they currently describe different experiments.
• Whether anyone runs flide verify on weights the authors did not test. This is the rare speed claim that ships its own falsifier: Apache-2.0 code, a published paper, 33 listed model families, and a subcommand whose whole job is to prove the forked answer matches the unforked one. The figure worth more than a speedup is a single exactness failure.
• Whether Google responds on autonomous retrieval specifically. Its help page already concedes human review of Connected Apps data and that a training sample "may just be the email or file". The unanswered question is narrower and entirely answerable: does Workspace retrieval fire without an explicit mention, and what fraction of a matched document enters context.
• Still open: Inkling 2 Small's Apache-2.0 weights, now callable but unreleased, flagged since the 6th. Voxtral 3 weights, flagged daily since 21 October, with nothing under mistralai since 15 October. No independent number yet for Tencent's DSpine drafter (7 November), Nanbeige4.3-33B (the 26th) or Ripple-65B Preview; no placebo control for the weight-space programming result (the 3rd); no published assessment from the three complexity theorists who agreed on 26 October to examine the Exponential Time Hypothesis manuscript; no account from Cursor of how the tampered extension reached its marketplace; and no fail-open rate from any decision-model vendor, asked here since 10 October and now thirty-one days.