The Frontier AI Wire is researched and drafted by Claude, an AI model made by Anthropic, under rules set by Attorney Jeffrey M. Beck. Every factual claim links to its source, with primary sources first. Where the brief goes beyond what a source says, it labels that as inference. Attorney Beck reviews and approves each edition before it is published. Errors are corrected in place, marked where they occurred, and logged. Nothing is changed silently. How the Wire is made.
The checking phase has begun, and it is not one queue. Anthropic posted its own refutation of the Exponential Time Hypothesis — found, it says, by an internal Claude model with no human input, and stated in Lean as “ETH is false” — days after OpenAI's. The physicists holding OpenAI's Yang–Mills proofs set out how they will check them, while a group of mathematicians says to release them now. Artificial Analysis moved most of its index onto private test sets and the gap between closed and open-weight models went from 3 points to 12, and Anthropic published a randomized trial in which analysts working with an agent took its secretly assigned side 86% of the time. No new corrections.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
Anthropic posts a second refutation of the Exponential Time Hypothesis, and its Lean file states the conclusion outright
A paper titled “A 2o(n) Time Algorithm for 3‑SAT,” with Anthropic as contributor, is posted at hexagonmath.org as version 1 dated 27 October, under CC BY 4.0 and labelled “primarily AI-generated text.” Its provenance note says the core algorithm was found on 16 October by an internal Claude research model, using about 14 million output tokens and no human input, and that researchers then improved the presentation. 16 October is also the date OpenAI gives for its own natural-language ETH proof (edition 033). The paper cites OpenAI's manuscript as concurrent work. Anthropic mathematician Levent Alpöge posted the accompanying repository at 00:00 ET on 27 October, inside edition 033's coverage window; that edition did not carry it.
The claim: deterministic 3‑SAT in 2o(n) time, which contradicts ETH, and counting (#3‑SAT) in 2O(n log log n/√log n) — the same as the sharper bound in OpenAI's manuscript. The method is tensor-network contraction. No outside expert has confirmed either lab's result. TechCrunch quotes MIT's Ryan Williams, one of the people OpenAI asked to check its version, as saying some parts are routine while “two or three” look “alien,” and that he does not think OpenAI's algorithm “could possibly be the best algorithm.” Scientific American's account is that neither result has been verified.
Some problems get explosively harder as they grow. The classic example, 3‑SAT, asks whether a long list of yes-or-no conditions can all be met at once, and every extra variable doubles the number of combinations to try. For decades, computer scientists have worked on the assumption — the Exponential Time Hypothesis — that no clever method can escape that doubling in the worst case. Two labs now say their models found a method that does.
Why it matters: ETH is a load-bearing assumption. Many results of the form “this problem cannot be solved much faster than that” are proved by assuming it, so if it falls, those results need revisiting, as Williams told TechCrunch. What it does not mean: this is not P vs NP, and it is not a practical shortcut — the claim concerns how running time grows as problems get very large. According to Williams, OpenAI checked and is “reasonably confident” its version breaks no cryptography.
The brief's read: the larger shift is where the bottleneck now sits. By the labs' own accounts, two AI systems reached the same claimed result within days, and the human work is checking it, not finding it.
What the Lean development states. Per the repository README, the file MainResults.lean states the main theorem and “ETH is false” using Mathlib's Turing-machine model. That is the step edition 033 could not find in OpenAI's Lean documentation, which stated a counting bound rather than ETH itself. The README says the only sorry warnings come from Challenge.lean, and that the main results depend only on the three standard axioms (propext, Classical.choice, Quot.sound). It is built on Lean/Mathlib v4.33.1, checked in CI with the Comparator tool and rechecked with the independent nanoda kernel. The Lean code was itself written by AI.
What the Lean development does not settle. The README says nothing machine-checks that the formal statements match the paper; an appendix and a PaperIndex.lean file are there for human review. A kernel check confirms that the formal theorem follows from its definitions, not that the definitions capture ETH as complexity theorists state it. That gap is where outside checking now sits for both labs.
A claim about the model, also Anthropic's own: the README says Claude Opus 5.5 at high effort can rederive the result from the paper's sketch. Nothing independent bears on that yet.
Sources The paper, v1 (primary) · anthropics/formal-math, eth (README and Lean) · @__alpoge__ — read 28 Oct from the Frontier Wire Sources list · TechCrunch, 27 Oct · Scientific American, via Yahoo
Artificial Analysis moves most of its index to private tests, and open-weight models fall 12 points behind
Artificial Analysis released Intelligence Index v5 late on 27 October. Its headline change: 60% of the index now rests on evaluations with private test sets or held-out answers, a majority for the first time. It launched a Progress Index v1 alongside, to track capability over time on one continuous scale. Claude Opus 5.5 leads at 58, and Anthropic and OpenAI hold the top five places.
| Model | v5 score | Cost per task |
|---|---|---|
| Claude Opus 5.5 (max) | 58 | $10.04 |
| GPT‑6 Astra (max) | 57 | $6.24 |
| GPT‑6.1 Sol (max) | 55 | $1.27 |
| Gemini 4 Argon (high) | 50 | — |
| Kimi K3.1 Pro (weights not yet released) | 46 | — |
| Claude Haiku 5.5 (max) | 36 | — |
| DeepSeek V4.1‑Flash‑1017 (max); GLM‑5.3 (max) | 32 | — |
The open-weight result is the finding. Kimi K3.1 Pro, the top open-weight model but with weights not yet released, trails the leader by 12 points, against 3 on v4.3. Among models whose weights are out, DeepSeek V4.1‑Flash‑1017 and GLM‑5.3 lead at 32, below Claude Haiku 5.5. The gap is widest on HLE‑Diamond and Terminal‑Bench‑Science, where Kimi scores 37% and 34% against 68% for Opus 5.5 and 63% for Astra. On cost, of the ten highest-scoring models shown, GPT‑6.1 Terra (max) is cheapest at $0.96 a task and Claude Fable 5.1 (max) most expensive at $13.04. Ethan Mollick called the open-weight finding one to “take with a grain of salt” that matches the general sense in the field; Nathan Lambert wrote that he felt “a need to defend the AAv5 index and open models at the same time.”
A benchmark only measures ability if the model hasn't effectively seen the test. Public test questions can end up in training data, and scores can then rise without the underlying skill improving. Moving most of the index onto questions that are kept private is meant to close that gap. On those questions, the best models anyone can download and run themselves fall further behind the leading closed models than the old index showed.
Two cautions. This is one organisation's index, and because the tests changed, a v5 score cannot be compared with a v4.3 score — the widening gap is a finding about the new tests, not proof that open models went backwards. The brief's read: the size of that gap is a running measure of how concentrated top capability is in a few labs.
What changed. New: FrontierCode 2.0, Cognition's private benchmark of real coding tasks graded on whether a maintainer would merge the change, and Terminal‑Bench‑Science, 70 agentic research tasks. Upgraded: HLE‑Diamond, a 1,000-question refinement of Humanity's Last Exam; AA‑Omniscience v1.1, with saturated and public questions removed; and CritPt v2, scored on a private 40-question hard subset (the 70 questions minus the 30 that seven frontier models solved in all five attempts at launch). Kept: AA‑Briefcase, AutomationBench‑AA, GDP.pdf, Terminal‑Bench 4.0. Removed: GDPval‑AA v2.1, SciCode, AA‑LCR v1.1. Weighting: Knowledge Work, Coding and Science are each 25%; Knowledge Work splits AA‑Briefcase 15, AutomationBench 5, GDP.pdf 5, and Coding splits FrontierCode 2.0 15, Terminal‑Bench 4.0 10.
No model leads everything. Opus 5.5 leads FrontierCode 2.0 (46%), HLE‑Diamond (68%) and AA‑Omniscience v1.1. Astra leads Terminal‑Bench‑Science (63%) and the CritPt hard subset (72%). Claude Sonnet 5.5 leads AA‑Briefcase (1812 Elo), Kimi K3.1 Pro leads Terminal‑Bench 4.0 (67%), Gemini 4 Argon leads AutomationBench‑AA (78%) and Muse Wonder Full leads GDP.pdf (33% all-pass).
Check the version before comparing. v5 scores are not on the v4.3 scale. Kimi K3.1 Pro's 55 (edition 032) and GPT‑6.1 Terra's 48 (edition 033) were v4.3 figures, and the same day Artificial Analysis still scored Ling 3.1 Flash 1019 at 44 on v4.3. As read at compile time, the artificialanalysis.ai homepage still showed v4.3.2. All figures are Artificial Analysis's own runs; none are lab-reported.
Sources @ArtificialAnlys, v5 thread — read 28 Oct from the Frontier Wire Sources list · @ArtificialAnlys, open-weight gap · @emollick · @natolambert · artificialanalysis.ai
Analysts working with an agent took its secretly assigned side 86% of the time
Anthropic's “How AI agents shape experts' judgment calls” (Herdağdelen, Joren, Tamkin; 27 October) is a randomized trial with recruited analysts working on a fictional farming cooperative's decision. The analysts used Claude Code with Opus 4.8 at high effort; the median session ran 65 minutes. On each of four judgment calls in the analysis, the agent's lean was set in secret by coin flip. Of 61 recruited, 11 declined to use the agent, leaving 50 analysed.
| Decision | Matched |
|---|---|
| Outcome measure | 92% |
| Weighting | 90% |
| Comparison method | 82% |
| Coverage | 81% |
| All four | 86% |
The leans moved the bottom line too: they predicted 82% of final recommendations, against a 61% baseline. Among analysts who stated a prior preference (n=38), 60% switched when the agent disagreed with it and 6% when it agreed (confidence interval on the gap, 24–82 points). 77% reused the agent's arguments and 15% credited it, and 52% rated the final recommendation as more their own. No measured trait predicted who deferred. The paper's framing: “Keeping a human in the loop assumes that the human's judgment provides an independent check.” Mollick called the result “important, but not unexpected.”
The caveats Anthropic states. The leans were deliberately amplified; it is one task, in one field, with one model. Serious methodological errors appeared in 6 of the 11 reports done without the agent against 3 of the 50 done with it, but the 11 chose not to use the agent, so that comparison is confounded rather than a measured benefit.
The baseline worth noticing. With no assigned lean, Opus 4.8 made the same five choices in 112 of 123 runs. An agent's default is not random, so in ordinary use the steer would run in a consistent direction rather than averaging out. That last step is the brief's inference from the two figures, not the paper's stated conclusion.
Sources Anthropic research (primary) · @emollick — read 28 Oct from the Frontier Wire Sources list
The physicists holding OpenAI's Yang–Mills proofs set out five steps before release; a mathematicians' group says release now
The Yang–Mills Working Group at UC Santa Barbara's Kavli Institute for Theoretical Physics — Michael Douglas, David Gross (chair), Arthur Jaffe, Igor Klebanov and Nikita Nekrasov — published a statement dated October 2026, posted by KITP on the evening of 27 October. It says OpenAI asked Gross on 16 October to take custody of the AI-generated material, and that the committee has “no responsibility to OpenAI.” Its five steps: check the definitions; check that the Lean matches the informal treatment; release the Lean; deepen understanding with experts; then release its assessment together with the full 1,147 pages of LLM-generated proofs. “Before we release it, we would like to be convinced that it is correct.” It gives no dates.
The statement also says that “confinement and the mass gap remain very important open challenges” — consistent with edition 033's note that no mass-gap claim was found on OpenAI's page, and so with this not being a claimed Millennium Prize solution. On the other side, the Advisory Group on Mathematics and Artificial Intelligence, whose nine listed members include De Lellis, Gowers, Hairer, Vakil and Witten, issued a recommendation dated 26 October for immediate release, saying there is “no scientific reason for delaying the release any further,” after a 25 October statement on the disadvantages of designating mathematicians to control releases of AI output.
Yang–Mills theory is the mathematics underneath much of the Standard Model of particle physics, including the strong force that binds quarks inside protons and neutrons. Physicists have used it with enormous success for half a century, but nobody has proved, to mathematicians' standards, that the four-dimensional theory is well defined. That proof, together with the “mass gap” — showing that its lightest particle must have positive mass — is one of the Clay Institute's $1 million Millennium Prize problems. OpenAI claims the first part; the committee says the mass gap remains open.
The brief's read: the dispute is about a question that did not exist until recently — when a machine produces 1,147 pages that no human has fully read, who decides when the world sees them? One side wants experts to vouch for the work first; the other wants the whole community checking at once.
Sources KITP Yang–Mills Working Group statement (primary, PDF) · @KITP_UCSB — read 28 Oct from the Frontier Wire Sources list · AGMAI statements (primary) · Scientific American, via Yahoo
Axiom says its agents formalised the classification of finite simple groups in Lean in 24 days, and reports two flawed steps in the published proof
Axiom's post, dated 26 October and announced on X on 27 October, says an agent swarm produced a sorry-free Lean 4 formalisation of the Classification of Finite Simple Groups — about 43.1 million lines — in 24 days, working through 108 milestones and about 380 tasks, and reusing about 1.2 million lines from the Qiuzhen‑CFSG project. It follows the first-generation proof. The GitHub repository states Lean v4.35.0‑rc3, only the three standard axioms, no native_decide, and the TauCeti statement of the theorem; a build of under 24 hours is planned. Lyons, Nakano and Solomon are named as external advisers.
Axiom lists 34 “noteworthy deviations” from the published proof, mostly minor. Two it singles out: a step in Walter's 1986 memoir, where Axiom gives PSL2(49) as a counterexample to an unjustified assertion that did not overturn the conclusion, with a new route set out in an accompanying paper; and a lemma in Aschbacher–Smith (2004), where an SL2(5) counterexample led to a corrected inequality. Asked by Runtime Wire whether anyone outside the company had rebuilt the proof, CEO Carina Hong said “Not yet,” and that Axiom is publishing build artifacts so others can.
Finite simple groups are the basic building blocks of symmetry: every finite group is built from them, much as every whole number is built from primes. The classification says exactly which building blocks exist — several infinite families plus 26 exceptions, the largest being the “Monster.” Its human proof runs to tens of thousands of journal pages by about a hundred authors over several decades, long enough that checking all of it is beyond any one person.
A Lean formalisation means a computer has checked every step, which matters most for exactly this kind of proof. Finding two flawed steps — with the theorem still standing — shows that even celebrated published proofs contain errors that formal checking catches. The caveat is the one in the headline: until someone outside Axiom rebuilds the 43 million lines, the check rests on Axiom's own run.
Sources Axiom (primary) · AxiomMath/CFSG on GitHub · @axiommathai · Runtime Wire, 27 Oct
Google DeepMind's Genie 4 adds native audio and live text steering to a real-time world model
Announced on 27 October, Genie 4 generates explorable environments in real time with physics and native audio — DeepMind calls it the first world model with native audio — and accepts text prompts that change the scene while it runs. DeepMind says it is built on Gemini Omni, with Gemini 3.8 Flash acting as a guiding agent. It is a research preview for trusted testers, with rollout planned to Google AI Ultra subscribers in Project Genie and to early API and AI Studio partners, and a waitlist. All output carries SynthID. DeepMind lists drift and object permanence as limitations. It published no quantitative specifications such as resolution, frame rate or session length, so there is nothing yet to check independently. Mollick, who has access, called it “glitchy, but also impossibly advanced compared to a couple years ago.”
Sources Google DeepMind (primary) · @demishassabis — read 28 Oct from the Frontier Wire Sources list · @emollick
Anthropic reports drug-discovery results with six partners, and buys a one-person biotech
“Early progress on accelerating drug development with Claude” (27 October) describes research partnerships with six organisations. Anthropic also says it acquired Mithril Medicine, a one-person company working on a rare muscular dystrophy, and is coordinating work across 11 labs; it reports going from idea to mouse results in under 18 months and says human trials could begin in late 2028. All figures below are Anthropic's and its partners'; most are preclinical.
| Partner | Reported result |
|---|---|
| Generate Biomedicines | A design about 70% shorter, binding about 10× the receptors |
| Divergence Labs | About 35,000 antibody designs in about a month; 20 dislodge IgE, 3 better than the standard drug |
| BigHat | About 1,010 variants in about 10 weeks; more than 10× tighter binding, +18°C stability, more than 100× slower clumping |
| Enveda | 1,061-page review: caught all 11 human-found issues plus 7 more; review time down 89%; a regulatory draft in 4 hours against about 3 months, with one minor error |
| TCG / Pleiades | 45 target–disease pairs |
| Bristol Myers Squibb | Clinical-trial planning; expansion planned for 2027 |
Sources Anthropic (primary) · @AnthropicAI — read 28 Oct from the Frontier Wire Sources list
Also on the wire
Confirmed, and not enough on their own to move the ranking.
-
Anthropic and GSA extend OneGov to state and local government at 40% off
On 27 October Anthropic said a new OneGov agreement with the General Services Administration offers Claude Enterprise, Claude for Government and the API to federal, state, local, tribal and territorial governments at 40% off, with no minimum spend. Anthropic puts the estimated savings at $1B and says about 500 organisations already use Claude; Claude for Government became generally available on 30 September with FedRAMP High.
Sources Anthropic (primary) · @AnthropicAI
-
Oracle will offer Google TPUs on its cloud
At Oracle AI World on 27 October, Oracle said it will deploy Google TPUs in OCI from 2027, which SiliconANGLE reports is the first time Google has supplied them to another hyperscaler. SiliconANGLE, citing CNBC, also reports Cerebras accelerators and OpenAI's Jalapeño inference chip in Oracle's plans, alongside ChatGPT and Codex integrations in Oracle's AI Workspace.
Source SiliconANGLE
-
CXMT reportedly starts HBM3E production at low yields
Reuters reported on 28 October, citing sources, that China's CXMT will begin limited HBM3E production within about two months despite yields of 20–30%, against roughly 70% at SK hynix, and has sent samples to Chinese chip designers; Seoul Economic Daily adds that performance is closer to HBM3. This is sourced reporting; no CXMT statement was found at compile time.
Sources Reuters · Seoul Economic Daily
-
Epoch AI maps where AI chips go
Epoch AI launched an AI Chip Imports explorer on 27 October. Its headline estimates: 72% of AI compute deployed since 2022 is in the United States and about 10% in China, and the US share of new compute rose from about 60% a year in 2022–24 to 72% in 2025 and 78% in the first half of 2026. The explorer page could not be fetched at compile time, so the figures are from Epoch's own post.
Sources @EpochAIResearch — read 28 Oct from the Frontier Wire Sources list · epoch.ai/data/ai-chip-imports
-
The standing disclosure pages, read this morning
OpenAI's misalignment page still prints 20 of 20 reports and three Notices; the newest entry was first posted and last updated 20 October. The Deployment Safety Hub's newest item is still the 26 October GPT‑6.1 Terra addendum (edition 033). Anthropic's coordinated vulnerability disclosure dashboard still serves its 2 October 19:47 UTC snapshot: 6,157 disclosed across 591 projects, 516 patched, 219 CVEs and 365 GitHub advisories. The Transparency Hub's Model Report is dated 2 October; no newer date was found there.
Sources OpenAI misalignment reports · OpenAI Deployment Safety Hub · Anthropic CVD dashboard · Anthropic Transparency Hub (all read 28 Oct)
-
The lab desks, as read this morning
anthropic.com/news's newest items were the 27 October GSA and drug-development posts; its research index added the drug-development and judgment-calls posts (items 03 and 07). No item dated 27 October was found on openai.com/news, whose newest was 26 October. mistral.ai/news's newest was 21 October. x.ai/news returned only older items as fetched, so nothing is asserted about it. ai.meta.com/blog's newest was still 27 July. DeepSeek's news link still pointed to its 17 October item. Qwen's Hugging Face organisation's newest model was about 15 days old. METR's newest was 22 October, Epoch AI's blog 16 October, ARC Prize's blog 3 September. DeepMind's index printed month-only dates; Genie 4 was dated from the post itself. These describe what those pages showed at compile time, not whether the organisations published elsewhere.
Sources anthropic.com/news · anthropic.com/research · openai.com/news · Mistral · x.ai/news · METR · Qwen on Hugging Face
Checked and spiked
Items that circulated but did not survive verification.
“Mistral opens the Timestral family.” AI Weekly's alerts for 28 October listed Mistral's Timestral time-series models as a new release. Mistral's own post, “Introducing Timestral,” is dated 15 October, and the release was covered in edition 026. The alert's paper link also pointed to a different paper, the one for the Hy‑Nomas world model. Nothing about Timestral is new this week.
Sources Mistral, 15 October · AI Weekly alert
Corrections
Errors in this brief — fixed in place above, logged here.
No new corrections. Editions 002, 003, 006, 008, 014, 015 and 025 carry corrections, archived with their editions. The standing rule after edition 015 — that an absence is reported as “not found at compile time, at this URL” or not at all — applies throughout, including the lab desks. Item 01 notes that Anthropic's ETH repository was posted inside edition 033's window and not carried there; that edition made no statement the posting contradicts, so it is recorded as an omission rather than a correction. On sourcing: the private source list was reachable at compile time through the in-app browser; the Artificial Analysis v5 results, the Epoch explorer, the KITP statement and the Alpöge posting came through it.