AI Daily Brief · Nov 3, 2026 | Anthropic's amended S-1 shows $16.96B in Q3 revenue; MIT has GPT-6 Astra edit a small model's weights directly; FT says Nvidia told clients to limit China business

Anthropic’s amended IPO filing dominated the past 24 hours. Q3 revenue came in at $16.96 billion, roughly half again as much as Q2, but the company swung from an operating profit to an operating loss, and a new line in the risk factors says internal estimates indicate future models may soon meet its Responsible Scaling Policy threshold for “dramatically accelerated progress attributable to AI.” On the research side, MIT CSAIL had GPT-6 Astra edit the weights of Qwen3.5-0.8B directly, changing only 0.016% of them and sharply raising its math scores, and third parties have already reproduced the result. OpenAI issued a run of announcements: a security sprint with Cathedral Systems and bigger bug bounties, former Army Secretary Dan Driscoll joining the company, the in-house Git backend Pando and a new benchmark, LifeVal. The FT reported exclusively that Nvidia told cloud clients to sell at least 51% of their compute to non-Chinese companies, and STAR Market Daily said DeepSeek plans to pretrain V5 on Huawei’s Ascend chips. Sam Altman said the internal model behind the Navier-Stokes result is the next GPT and will ship soon, fueling unverified GPT-6.2 rumors.

Today’s highlights

• Anthropic’s S-1/A: Q3 revenue of $16.96B, an operating loss of $567M and a net loss of $23.29B, plus about $895.4B in compute-related commitments; the filing says future models may soon reach its RSP threshold for AI-accelerated progress [Company claim].
• MIT CSAIL used GPT-6 Astra to change 131,072 weights in Qwen3.5-0.8B, lifting MATH-500 from 7.9% to 48.1%; a Hugging Face re-run got 51.0% [Claimed in the paper] [Not peer-reviewed].
• OpenAI’s bug bounty now pays up to $172,000 including bonuses, and it is running an internal security sprint with Cathedral Systems; its in-house Git backend Pando cut median push latency by nearly 80% [Company claim].
• FT: Nvidia verbally told cloud clients that at least 51% of future services must go to non-Chinese companies [Media exclusive]; STAR Market Daily says DeepSeek V5 will be pretrained on Huawei Ascend chips [Media report, not confirmed by the company].
• Anthropic launched a public beta of message threads in the Messages API, cutting median time to first token on long Claude Code requests by about 35% [Company claim].

1. Anthropic’s amended S-1: $16.96B in Q3 revenue, and internal estimates say it may “imminently” reach its RSP threshold for AI-accelerated R&D

The numbers: According to the S-1/A on SEC EDGAR, Anthropic’s revenue for the quarter ended Sept. 30 was $16.956 billion, up from $11.555 billion in Q2 and $1.382 billion a year earlier. It posted an operating loss of $567 million after an operating profit of $974 million in Q2. Adjusted operating income was $2.463 billion, an adjusted operating margin of 15% (19% in Q2), and the net loss was $23.291 billion. Nine-month revenue was $33.242 billion and the net loss $104.039 billion, of which $94.324 billion came from fair-value changes in convertible notes, derivative liabilities and SAFEs; the filing attributes this mainly to convertible notes rising in value as the company’s valuation increased.

Compute commitments: As of and after Sept. 30, purchase commitments and lease agreements totaled about $895.4 billion, substantially all related to compute and infrastructure. About $447.5 billion is expected to be incurred before the end of 2030, with substantially all of the rest between 2031 and 2045.

The RSP language [Company claim]: A new sentence in the risk factors says Anthropic’s internal estimates and its expectations about future models, rather than current ones, indicate it may imminently meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” It has not designated any current model as exceeding that threshold and is still assessing the underlying data. The filing also concedes that it has in the past not identified, and may in future not identify, such an acceleration until after it has happened.

What the threshold means [Opinion]: Citing the RSP’s definition, @deredleritt3r said it roughly means AI capabilities progressing at twice the rate expected without AI help, compressing two years of progress into one (369 likes). @AndrewCurran_ said the passage is on page 29 of the S-1/A (926 likes).

Reactions: @MTSlive posted a flash (2,499 likes, about 640K views). @unusual_whales listed the Q3 figures (4,356 likes, about 1.15M views) and followed up with the nine-month loss and compute commitments (2,358 likes). @kimmonismus relayed the RSP passage (1,872 likes).

Sources: @MTSlive 1 · @kimmonismus 1 · @deredleritt3r · @AndrewCurran_ 1 · @unusual_whales 1 · @unusual_whales 2 · SEC EDGAR: Anthropic S-1/A · CNBC · Reuters (via Investing.com)

2. MIT CSAIL had GPT-6 Astra edit Qwen3.5-0.8B’s weights directly, changing just 0.016% and lifting its math score from 7.9% to 48.1% [Claimed in the paper] [Not peer-reviewed]

How it works: Authors Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab call the approach weight-space programming. According to their blog, GPT-6 Astra did no gradient training. Instead it located a single direction associated with the small model’s reasoning loops, in the MLPs of layers 15 and 19, and subtracted it from 64 output columns in each, changing 131,072 weights, or 0.016%.

Results: MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval from 21.0% to 62.3%, while the share of runs hitting the token limit fell from 93.0% to 33.1%. The gains mostly come from suppressing repetitive reasoning loops, not from adding knowledge.

Replication [Third-party claim]: The authors said the edited model is now on Hugging Face, and that @akhaliq rebuilt it from the published logs and re-ran MATH-500, getting 51.0% (292 likes). @AlexKrentsel said he had reproduced the blog’s results and argued this is a new capability that has emerged only in the latest models [Opinion] (6,727 likes).

Reactions: @rikiyatakehi’s announcement drew 50,964 likes and about 6.93M views, the most of any AI post in this issue. @kimmonismus said it points toward AI diagnosing and improving other models directly (5,514 likes). @emollick wrote “Waiting for replication but this is a pretty big deal” [Opinion] (1,419 likes). @op7418 posted an explainer in Chinese (74 likes).

Sources: @rikiyatakehi 1 · @emollick 1 · @kimmonismus 2 · @AlexKrentsel · @op7418 · @rikiyatakehi 2 · Authors’ blog · Hugging Face model · VentureBeat

3. OpenAI on security: an internal red-team sprint with Cathedral Systems, bigger bug bounties and former Army Secretary Driscoll on board

Security sprint: OpenAI said it is planning its next internal security sprint, in which researchers from Cathedral Systems will work alongside its teams, using OpenAI’s models to test its defenses. @AndrewCurran_ noted that Cathedral is a cybersecurity startup founded by former DOGE staffers that recently hacked Anthropic’s Slack (285 likes).

Higher bounties: According to OpenAI, the midpoints of reward ranges for medium, high and critical findings more than quadrupled on average. Critical findings in core products now pay up to $30,000, exceptional findings up to $100,000, and up to $172,000 with bonuses. The new ranges also apply to reports awaiting a bounty, and for bounties awarded in the past 30 days OpenAI will pay the difference if a finding qualifies for more. It also announced a PROMPT2PWN event in Oakland on Nov. 20–21. @OpenAI’s announcement drew 2,911 likes.

National security role: Dan Driscoll, the 26th U.S. Secretary of the Army, has joined OpenAI in the newly created role of Senior Counselor for National Security Modernization. @AndrewCurran_ (158 likes) and @Reuters reported the move.

Sources: @AndrewCurran_ 2 · @OpenAI 1 · @OpenAI 2 · @AndrewCurran_ 3 · @Reuters · OpenAI: bigger bug bounties · OpenAI: next security sprint · OpenAI: Dan Driscoll joins

4. OpenAI engineering and evals: in-house Git backend Pando cuts push latency by about 80%; new benchmark LifeVal tests everyday tasks [Company claim]

Pando: OpenAI built its own Git backend for its internal monorepo, and engineers and agents keep using Git as before. The company says median push latency fell by nearly 80% and that Pando now serves nearly a petabyte of Git data a day, peaking at 57TB in one hour. It was written in Rust with Codex in 10 weeks and has been in production since April, and the team also contributed fixes back to Git. @gdb said OpenAI is considering offering it to others and invited interested parties to email him (1,303 likes).

LifeVal: A benchmark for how well a personal assistant handles everyday tasks, designed from the American Time Use Survey, the Multinational Time Use Study and a representative survey of 3,995 people. It covers caregiving, health, education, household administration, personal finances, shopping and government services. OpenAI says it has used it internally for a year to improve its models; the benchmark itself has not been released.

Reactions: @OpenAIDevs’s Pando announcement drew 5,482 likes and about 1.02M views. @thsottiaux shared it, saying the team knows some Git-fu (3,008 likes). @OpenAI’s LifeVal post drew 6,117 likes.

Sources: @OpenAI 3 · @OpenAIDevs 1 · @OpenAIDevs 2 · @thsottiaux 1 · @gdb · OpenAI: Pando · OpenAI: LifeVal

5. FT: Nvidia tells cloud clients to sell at least 51% of compute to non-Chinese companies [Media exclusive]; DeepSeek V5 reportedly to be pretrained on Ascend [Media report, not confirmed by the company]

Nvidia: Citing five people familiar with the matter, the FT reported exclusively that Nvidia has told cloud and data-center clients that at least 51% of their services must go to non-Chinese companies from now on. The requirement was delivered verbally with no written documents, applies to clients worldwide and to future contracts, and does not affect contracts already signed. Nvidia neither confirmed nor denied it and issued a statement; Chinese foreign ministry spokeswoman Mao Ning said she was not aware of the matter. @FT (130 likes) and reporter @zijing_wu (138 likes) posted the story.

DeepSeek V5: Citing several supply-chain sources, STAR Market Daily reported that DeepSeek plans to pretrain its next model, V5, on Huawei Ascend chips; Ascend had mainly been used for inference before. DeepSeek did not respond. @jukan05 shared the report (267 likes). @poezhao0605 noted that the V4 technical report does not say which chips were used for training and said he would wait for the V5 report (5 likes). @kimmonismus called it China’s next step toward independence from U.S. chips [Opinion] (360 likes).

Sources: @FT · @zijing_wu · @jukan05 · @kimmonismus 3 · @poezhao0605 · FT · Malay Mail (relaying the FT) · Investing.com · STAR Market Daily

6. Altman says the internal model that cracked Navier-Stokes is the next GPT and is coming soon [One-sided claim]; GPT-6.2 rumors heat up [Unverified]

Altman’s claim [One-sided claim]: According to The National, Altman said at ADIPEC in Abu Dhabi on Nov. 2 that the internal model which tackled the Navier-Stokes problem, running as about 10,000 agents for about four days straight, is the next GPT and will be released soon. He gave no date. The math result has not been recognized by the Clay Mathematics Institute. @chetaslua (1,264 likes), @kimmonismus (1,740 likes) and @MTSlive (2,028 likes) relayed the remarks.

Rumors [Unverified]: OpenAI’s @thsottiaux posted just “Le GPT Fat” (7,811 likes, about 1.23M views), widely read as a teaser for a new model. @chetaslua claimed the new model’s lowest setting beats the current Astra at max (1,094 likes). @kimmonismus said Astra already appears to be routing some users to GPT-6.2 (940 likes). @TokenGremlin said Dots will come to Plus users powered by GPT-6.1 Terra and labeled it confirmed, without citing a source (909 likes). As of this writing, OpenAI had made no announcement.

Sources: @TokenGremlin · @chetaslua 1 · @kimmonismus 4 · @MTSlive 2 · @thsottiaux 2 · @chetaslua 2 · @kimmonismus 5 · Snape Report (citing The National)

7. Three from Anthropic: message threads in public beta, a research residency and FlashProve, a provable-inference prototype [Company claim]

Message threads: The Messages API now has a public beta in which the server caches the conversation between turns, so clients send only new messages and the previous message ID. Anthropic says that in Claude Code, median time to first token on requests of 500K to 1M tokens is about 35% lower, and the median request in long conversations shrank from about 800KB to 1.8KB. It is billed at standard token rates and available on the Claude Platform, AWS, Microsoft Foundry and Google Cloud. @ClaudeDevs’s announcement drew 7,421 likes.

Anthropic Residency: A paid, fixed-term research program for academics, technical founders and staff-level engineers, with no ML background required. Placements run six weeks, three months or six months, pay is equivalent to $350K–$500K a year, and applications close Nov. 22. @AnthropicAI’s post drew 31,214 likes and about 9.59M views.

FlashProve: A research prototype from Anthropic’s Frontier Red Team that proves a given output was produced by a specified model, not a distilled or tampered version, without revealing the weights. In a demo on Llama 4 Maverick (400B), one H100 proves about 110 tokens per second, each 1,024-token request yields a proof of about 19MB, and GPU time is 70–95x that of vLLM inference. It is open source under Apache 2.0. @scaling01 shared the research (309 likes).

Sources: @ClaudeDevs 1 · @ClaudeDevs 2 · @ClaudeDevs 3 · @AnthropicAI · @scaling01 · Claude: message threads · Anthropic Residency · Anthropic: provable inference

In brief

• Mistral Large 4 technical report [Company claim]: Following the open-weight release on Oct. 29, Mistral published a technical report on alphaXiv. It lists 1T total and 52B active parameters, a FrontierCode 2.0 score of 19.7% and a Terminal-Bench 4.0 score of 41%; the NVFP4 checkpoint is 606GB and fits on four GB300s. The report also says 16x KV-cache compression hurt agentic performance, so the team performed model surgery and switched to DeepSeek Sparse Attention (@PierreStock, 519 likes; @MistralAI, 311 likes; alphaXiv).
• Tutor Intelligence raises $75M: The robotics company closed a Series B led by a16z, with USV and Neo participating. Its robots are guided by remote human “tutors,” and that work trains its robot foundation models. The company says it is in almost four times as many sites as a year ago and expects more than 1,000 robots under contract by the end of 2027 [Company claim] (@joshgruenstein, 411 likes; @a16z, 224 likes; Business Wire).
• The shelf life of LLM findings [Not peer-reviewed]: Stanford’s Owen Queen and James Zou re-tested 50 widely cited LLM findings on 20 models. 44% no longer hold in today’s models, 34% have weakened substantially, 16% persist and 6% have strengthened. Capability failures such as letter counting and fabricated citations fade fast, while preference-type tendencies such as flattering users and judges favoring their own answers last longer (@james_y_zou, 831 likes; @emollick 2, 30 likes; project page).
• Hume Audio Decision API [Company claim]: Send audio plus plain-language questions (how many speakers, whether a caller is frustrated) and get probabilities back. Hume says it was trained on more than 10 million human judgments; access is by request for now (@hume_ai, 1,778 likes; Hume blog).
• Musk on cooling in space [One-sided claim]: Weighing in on space data centers, Musk said an aluminum radiator pointed away from the sun can easily reach about 1,300 W/m², while solar panels generate about 250 W/m², so the panels should have about five times the radiator’s area (@elonmusk 1, 13,140 likes). In the same discussion, @MillionInt posted “Space datacenters actually make sense” [Opinion] (@MillionInt, 1,924 likes).
• $200 plans are enough [Opinion]: @trq212 said that after a week, his takeaway is that you don’t need unlimited tokens to do great work, that the $200 plans are amazing and that this is the most expensive intelligence will ever be (@trq212, 7,415 likes).
• Grok Bot improvements: Musk posted “Grok @Bot improvements” with a link (@elonmusk 2, 6,610 likes).
• A map site built with Opus 5.5 [Personal test]: @emollick had Opus 5.5 build a site showing the fastest routes into every U.S. city and the fairest places to meet up, among other things (@emollick 3, 791 likes).
• A step toward recursive self-improvement? [Opinion]: @kimmonismus called weight-space programming a giant step toward recursive self-improvement (@kimmonismus 6, 721 likes).

---

About this column and how we verify

1. AI-assisted: AI Daily Brief is compiled with AI assistance; this issue was produced with Grok Bot and covers Nov 3, 2026 00:00 to 22:07 Beijing time. Anthropic’s S-1/A was accepted by the SEC at 22:24 Beijing time on Nov 2, after the previous issue’s cutoff, so it is covered here.
2. Every post checked: Each quoted post was read individually to confirm it exists, the author matches and it was posted within this issue’s time range, and all of them were cross-checked a second time through the X API. Like and view counts were read at about 22:15 Beijing time on Nov 3.
3. X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts come from the original posts, official pages and media reports, and the primary source wins in any conflict. Anthropic’s financials and RSP wording were checked directly against the filing on SEC EDGAR.
4. Items involving Grok Bot: The brief item “Grok Bot improvements” concerns Grok Bot itself, which helps produce this column; it was included on popularity alone, under the same standards as everything else.
5. Labels: [Company claim] = from a company announcement, filing or official account, not independently reproduced; [Claimed in the paper] = results reported by the researchers themselves; [Not peer-reviewed] = not yet peer-reviewed; [Third-party claim] = a replication or test reported by a third party; [Media exclusive] = an exclusive from a single outlet; [Media report, not confirmed by the company] = reported by media, not yet confirmed by the company involved; [One-sided claim] = one party’s account with no other corroboration; [Unverified] = circulating but not confirmed by an official or authoritative primary source; [Opinion] = personal judgment, not a statement of fact; [Personal test] = one person’s trial, not a systematic evaluation.