AI Brief, 7 November 2026: twice as fast, and most of the evidence redacted

The most consequential document of the window is an outside auditor's report on how fast a frontier lab is now moving, and the second most consequential thing about it is how little of the evidence the public is allowed to see. METR published an independent assessment of AI R&D acceleration at Anthropic on 5 November, late for this series and covered here for the first time. Four METR staff spent eleven weeks, 12 August to 29 October, with access to internal benchmark scores, unreleased model snapshots, training-compute metrics and about fifteen staff interviews. Their central estimate is that Anthropic's benchmark progress has accelerated 1.9 times since October 2025, meaning roughly two years of the prior rate compressed into one. Their headline caveat, placed above the executive summary, is that Anthropic exercised contractual redaction rights and removed "a majority of the evidence we found most informative" for that conclusion, and that the report therefore "does not achieve 'transparency about evidence'."

Anthropic has said something compatible, in a filing rather than a blog post. Its S-1 amendment of 2 November states that internal estimates and expectations about future models "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'," while noting that no current model has been designated as crossing it. The same filing concedes that much of its evidence comes from lagging indicators, and that the company "may in the future not identify a significant acceleration in our rate of development until after it has occurred."

What the window added on top was interpretation rather than data, and the interpretation travelled far better than the measurement. METR's report drew 44 points on Hacker News. A post arguing that the report implies very short timelines drew 238. A sceptical note about whether machine research is any good drew 184. The one artifact of the window that can actually be checked line by line, a machine-found proof of a conjecture open since 2008, with a Lean formalisation anyone can compile, drew ten.

• METR's independently estimated figure is 1.9x [95% CI 1.5–2.8x]; across methodological variants it ranges 1.1x to 2.9x. Anthropic's own published figure, in the Haiku 5.5 system card, was 1.7x.
• Anthropic staff self-report about 12x personal speedup from the latest internal model, in Slack polls METR calls "only very loose evidence". Uplift and acceleration are different quantities.
• Anthropic cancelled the contracted component that would have verified how compute is split between capabilities and safety work, midway through the engagement.
• METR notes that Claude wrote much of the material it was given and performed "the majority of our own analysis" — the audit of AI-driven acceleration was itself largely AI-conducted.
• PSI reports a research system running a non-public Anthropic model found a proof of the Aaronson–Ambainis conjecture on day three of a run; the Lean 4 formalisation is Apache-2.0 and public.
• Nous Research published Arete-27B, the only new model from twenty-one major organisations inside the window. Its propaganda rate falls from 22% to 0.9%; its factual score stays mid-table.

What a 1.9x acceleration is, and what it is not #

METR's number is not a productivity measurement. It is the ratio of two fitted trends in AECI, Anthropic's version of the Epoch Capabilities Index, computed with Epoch's public code: an ordinary least-squares piecewise-linear fit with a single estimated breakpoint, which landed on 7 October 2025. Confidence intervals come from bootstrapping benchmark scores stratified by model, and cover uncertainty in the scores but not uncertainty over which models enter the fit. Four never-released internal models are in it, dated by when most Anthropic staff could use the snapshot rather than by public launch, and that choice is most of the gap between 1.9x and Anthropic's own 1.7x. The honest summary of the spread is that the method moves the answer about as much as the finding does: Anthropic's own system card said 1.7x, METR's central estimate is 1.9x, and simply changing the link function in the same fit spans 1.1x to 2.9x.

The distinction that most of the day's commentary blurred is between uplift and acceleration, and METR spells it out. Let $r_0$ be the annual growth factor of a capability measure before the breakpoint and $r_1$ the factor after it. Acceleration is defined so that a value of 2 means two years of old progress in one year:

$$a=\frac{\ln r_1}{\ln r_0}$$

METR's own worked case: if a quantity grew $3\times$ per year, then doing two years of that in one year requires $3^2=9\times$ per year, and $a=\ln 9/\ln 3=2$ . A doubling of the rate is a far stronger claim than a doubling of anyone's throughput.

That is why the 12x figure cannot be read as twelve-fold acceleration. A one-off speedup is a level; acceleration is a change in a rate, and sustaining a faster rate against diminishing returns needs uplift that keeps growing. Ajeya Cotra's post of 6 November works the chain backwards: for progress to double through software alone, growth in compute multipliers must run perhaps 2.5 to 3 times faster; under a Cobb-Douglas model in which experiment compute and researcher labour matter about equally, required labour uplift goes as roughly the square of that, giving 4 to 9 times; allowing for months of already-accelerated progress plucking the easy results, she puts the figure needed today near 10x, and notes it lines up with the 12x staff self-report.

Three further disclosures deserve more attention than they got. Anthropic changed the scope midway to drop compute verification, the component that would have checked the split between capabilities and safety compute, before METR reached conclusions. METR had no access to sample from non-public models, worked on Anthropic-supplied laptops, and took no payment in either direction. And METR records that Claude was deeply involved in the majority of the data collection and analysis it received, and performed the majority of METR's own analysis: the audit of AI-driven acceleration was itself largely AI-conducted. A companion report published the same day concludes that Anthropic's public August risk report "overall understates the level of acceleration that was occurring."

A proof that compiles #

Against all of that, one item in the window is checkable without trusting anybody. PSI reported on 6 November that a research system it built internally, running instances of a non-public Anthropic model, found a proof of the Aaronson–Ambainis conjecture, open since 2008. The system had a complete proof on the third day of the run, and Lean, a proof assistant that mechanically checks every inference, accepted it. Set to related problems afterwards, the system improved the bound twice.

The conjecture settles when quantum computers can be dramatically faster. Shor's factoring algorithm wins because its input is promised to hide a repeating pattern; the open question was whether a quantum computer can beat a classical one exponentially on unstructured inputs. After $T$ queries a quantum algorithm's acceptance probability is a polynomial $p$ of degree at most $2T$ in the input bits, bounded in $[0,1]$ . Write the influence of bit $i$ as the mean squared change when that bit flips,

$${\mathrm{Inf}}_i[p]={\mathbb{E}}_X[(p(X)-p(X^{\oplus i}){)}^2]$$

where $X^{\oplus i}$ is $X$ with bit $i$ flipped. The conjecture says that if the variance of $p$ is at least $\epsilon$ , some bit must have influence at least polynomial in $\epsilon /d$ , with $d$ the degree. The repository proves

$${\mathrm{Inf}}_i[p]\ge \frac{8\sqrt{15}}{125{\pi }^4}\cdot \frac{{\epsilon }^{5/2}}{d^6}\ge {\left(\frac{\epsilon }{d}\right)}^7$$

stronger than the $(\epsilon /d{)}^{O(1)}$ the original asked for. An influential bit is exactly what a classical algorithm needs to imitate the quantum one by querying a few high-influence positions, so the theorem closes the gap: no exponential quantum speedup on unstructured yes-or-no questions. The result is about query complexity, the number of times an algorithm reads its input, not total runtime, and it does not say quantum computers are useless on unstructured problems, only that they cannot win exponentially on reads.

The engineering around the claim is the part worth copying. The repository is Apache-2.0 and holds 95 Lean files against Lean 4.33.0 and the matching Mathlib, with CI that rebuilds the formalisation on every push. The conjecture statement is isolated in Conjecture.lean, which imports only Mathlib and no project definitions, so a reader can check that the theorem proved is the theorem claimed without reading the proof:

theorem aaronson_ambainis : PrintedConjecture
theorem aaronson_ambainis_explicit : ExplicitBound

A fourth theorem is stated conditionally on a result of Agarwal and Ben-David that is deliberately not formalised, which is the right way to flag a dependency. Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan, the Stanford group that proved a special case in August, have read the proof and are writing an exposition; PSI quotes them calling it "a huge result". That is substantive, named corroboration, though not a referee report, and Scott Aaronson has not commented publicly.

The counterweight arrived separately and is not a rebuttal of any of the above. Arjun Panickssery posted a short note on LessWrong late on 5 November, at 132 karma, reporting that on three of his own side projects a model's research judgment is "very poor in a basic way": it misses the obviously important paper, runs experiments whose training sets are too small to separate the architectures being compared, and pursues trivial directions. His concession matches METR's: the uplift is real on hill-climbing tasks with fast feedback, where he got a 10x improvement on an enumeration problem by repeated prompting, though he credits the conceptual leaps to himself. He tested a non-Claude model, the projects are hobby-scale, and he never mentions METR. What makes the note useful is that it names the regime boundary METR also flags: the 1.9x is measured on well-specified, auto-gradable tasks, and METR says separately that it is more uncertain about broader work.

The only model of the day, and a number its own table undercuts #

Across twenty-one organisations that ship weights regularly, exactly one created a model repository in the window. Nous Research published Arete-27B at 19:31 UTC on 6 November under Apache 2.0: Alibaba's Qwen3.8-27B, supervised fine-tuned on about 52.8 million tokens at 64k context over 1,202 steps, to change how it answers politically sensitive questions. No RL, no preference optimisation, no architectural change. The teacher was Inkling, chosen partly because an open teacher avoids the terms-of-use problem that distilling a closed one would create. The 803-prompt evaluation is published as Propaganda_Eval under MIT, with the judge prompt and scoring scripts, and the training pairs are released too.

The reported movement is large. Answers containing propaganda fall from 22% to 0.9% and the fact score rises from 2.32 to 2.98 out of 5, while capability holds: Hermes Bench moves 65.77% to 65.60%, inside noise, and Nous declines credit for a Terminal-Bench-2.1 reading that rose from 67% to 70%.

The chart shows what the announcement does not. On factual recall Arete-27B lands sixth of nine, below GLM 5.3 and Muse Glimmer 30B and well below the Inkling teacher at 3.96. What the training bought is almost entirely the propaganda column, and the scoring rubric explains why: a single opposed fact scores the whole answer 0 regardless of everything else it got right. The metric is therefore closer to a framing detector than to a recall measure.

The published zero counts make that precise. If the 359 and 216 zero-scoring answers Nous reports correspond to the same runs as the 2.32 and 2.98 means, then the mean among answers that avoided a zero is $2.32\times 803/444\approx 4.20$ for the base model and $2.98\times 803/587\approx 4.08$ for Arete-27B. On prompts where neither model opposes a fact, the untrained Qwen scores marginally higher. The gain is in moving answers out of the zero bucket, a real and useful change, and a different claim from teaching the model more facts.

The caveats are load-bearing. Every figure is Nous's own, the judge is Grok 4.7 developed against a 25-prompt hand-graded pilot with the rubric revised three times, and the card concedes decoding settings were not identical across rows: Arete and Qwen ran at temperature 0, the comparison models at 0.6. The blog and the model card also disagree with each other, giving the propaganda rate as 1% and 0.9% and the training set as both 28,886 examples and "roughly fifty thousand samples". More structurally, the teacher was selected for scoring well on the same evaluation the result is reported against; overlapping prompts were removed, but choosing a teacher by the metric is close kin to training on it. And the post-training was text-only with the vision encoder frozen, so a model that accepts images and video has had none of those paths touched. The download counter reading zero hours after publication is counter lag, not evidence nobody tried it; the 101 likes arrived the same evening, and there is no Hacker News thread.

Corrections #

Ripple-65B Preview had an independent score before yesterday's brief said it had none. Plotly published its own run on 5 November, through OpenRouter on the harness it uses for every other model: 35 of 43 on its data-analysis exam at $2.21, third among US open-weight models there and behind Kimi K3.1 Pro at 42. Plotly cautions that one run is a single sample. The sharpest thing in it is not the score: on one question the model produced a table of twelve figures it had never computed, wrong by a factor of two to three, with nothing marking them as estimates. Ricursive's agentic-coding claims, including 80.52 on Terminal-Bench 2.1, remain self-reported, and that figure is on an older benchmark version than the one the official board now scores.

Voxtral 3's "end-of-October weights commitment", counted here since 21 October, is not traceable to Mistral's announcement, which offers only API access and self-deployment and says nothing about weights, a licence or a date. No Voxtral 3 weights exist and nothing has appeared under mistralai since 15 October, both still worth noting, but the overdue-promise framing was wrong and the day count is withdrawn.

Also notable #

• Gemini Omni 1.2 Flash, a video model, went generally available on 5 November: clips to 15 seconds, up from 10, and a 30% API price cut to $0.07 per second of 720p. Two things the announcement omits. Frame interpolation is capped at 10 seconds, so the headline duration and keyframe control cannot be combined. And while the API fell 30%, a filmmaker in the Hacker News thread reports the cost in Google Flow, where creatives actually work, rising from 10 to 18 credits for a six-second clip. Every published figure is Google's own preference win-rate against its own predecessor.
• The loudest item of the window, at 1,561 points, is a privacy report against Claude Code opened on 5 November: the model, given the user's account email in context for authorship purposes, pastes it into outbound User-Agent headers when it writes its own curl calls, because sites ask automated clients to identify themselves. A team member said on the 6th that the address will be removed from context early next week; that release has not shipped. The interim mitigation, a default-on privacy mod, is reported not firing on 2.1.317, 2.1.318 and 2.1.319; its own README calls it "a courtesy, not a boundary". A report of the same behaviour has been open since July.
• Yuyao Technology's MatterSeek, a materials model, has 2,915 likes on a Hugging Face repository containing no weights at all: images, two promotional videos, a logo and a proprietary licence. Its card says no public benchmark covers its variables, so "cross-model peer comparison is not applicable", and gives no numbers. The like count is the only measurable thing about it.
• Zyphra ZYFA, open FP8 attention kernels for AMD's Instinct MI355X, claims up to 1.76x over AMD's AITER. Zyphra's own figure, and the detailed comparison is not publicly retrievable, so it is unaudited.
• An independent harness leaderboard updated on the 6th answers the question this brief raised on the 5th, when a harness change moved one model 9.3 points. Running nine coding harnesses over the same models at three attempts each, it finds swapping the harness flips 29.1% of tasks while re-running the same harness flips 23.4%. The harness matters, but barely more than run-to-run noise, and its top five are statistically tied. Best to worst on one model is still 0.51 to 0.27, and the leader changes with the model.
• OpenAI reportedly closed about $11B at a $1.25T valuation, while Anthropic is reported to be targeting a listing before Thanksgiving. Both sourced to Bloomberg and the FT, so both are secondhand.
• transformers 5.22.2 shipped on the 6th, hardening Jinja sandboxing in chat templates. No vLLM, SGLang, Ollama or llama.cpp release landed in the window.

What to watch #

• Whether the redacted evidence ever becomes reviewable. The disclosure worth wanting is not the raw data but the per-stage acceleration estimates and the compute verification Anthropic cancelled, since the capabilities-versus-safety split is the one figure in the engagement that was contracted, dropped and never replaced.
• Whether anyone compiles the Aaronson–Ambainis formalisation who did not write it. This is the rare claim where independent verification costs a machine an afternoon rather than a referee six months: the repository pins Lean 4.33.0 and Mathlib, Conjecture.lean is readable on its own, and CI already builds it. A comment from Aaronson or Ambainis, or the Stanford group's promised exposition, would settle the mathematics; neither has appeared yet.
• Whether a third party runs Propaganda_Eval. The evaluation, judge prompt and scoring code are MIT and the weights Apache-2.0, so the decisive experiment is to rerun every row at matched decoding settings with a judge other than Grok 4.7. The figure worth more than the propaganda rate is factual recall among answers that opposed nothing, where the base model currently looks no worse.
• Still open: Inkling 2 Small's Apache-2.0 weights, flagged here yesterday, fall due the week of 9 to 15 November on Thinking Machines' own wording, so they are not yet late, though Artificial Analysis describes the release as date unconfirmed. The placebo control for the weight-space programming result, asked for on the 3rd, has not been run. Nanbeige4.3-33B still has no independent score, MiniMax no unquantised M3.1 Flash reference, Mage-VL-Cowork no independent OSWorld-Verified number, and Cursor no account of how the tampered extension reached its marketplace, now nine days. And no decision-model vendor has published a fail-open rate, asked here since 10 October and now twenty-eight days.