Anthropic Books a $17 Billion Quarter and Warns It May Soon Meet an AI Progress Threshold as OpenAI Quadruples Bug Bounties and Nvidia Caps Chinese Cloud Buyers

Anthropic’s Amended S-1 Shows a $17 Billion Quarter, $895 Billion in Compute Commitments and a New Warning About AI-Driven Progress

Anthropic filled in the most-requested blank in its IPO paperwork on Monday. Anthropic’s amended Form S-1 registration statement on SEC EDGAR, accepted at 6:24 AM Pacific on November 2, adds unaudited results for the third quarter. Revenue for the three months ended September 30 was $16.956 billion, up from $11.555 billion in the second quarter and $1.382 billion a year earlier. Revenue for the first nine months of 2026 reached $33.242 billion, against $2.545 billion in the same period of 2025. Compute and infrastructure expenses rose to $9.863 billion in the quarter, and the company swung to a $567 million operating loss after reporting $974 million of operating income in the second quarter. On the adjusted measure Anthropic prefers, it still reported $2.463 billion of adjusted operating income, a 15% adjusted margin.

The filing also explains some of the slowdown. Anthropic says higher usage was partly offset by a decrease of about $600 million from price reductions introduced with new model versions and contractual discounts between the second and third quarters. CNBC’s report on the revised prospectus notes that quarter-over-quarter growth fell to about 47%, from 144% in the second quarter and 132% in the first, that the third-quarter net loss was $23.29 billion, and that most of the $104 billion nine-month net loss came from changes in the fair value of convertible notes and derivative liabilities. The bigger number is the forward bill: Anthropic says that as of and subsequent to September 30 it has purchase commitments and leases with an aggregate committed spend of about $895.4 billion, substantially all for compute and infrastructure, including about $447.5 billion expected before the end of 2030.

Two new risk disclosures deserve attention. The first concerns the supply chain behind that compute, which the company describes as complex multi-party structures where chips are procured through cloud providers, leased by financing entities and run by other operators. The second is about the models themselves. In language that does not appear in the original October 26 filing, Anthropic says its internal estimates suggest it may soon meet one of its own Responsible Scaling Policy thresholds, while stressing that it has not designated any current model as exceeding it.

“…certain of our compute providers and infrastructure partners depend on third-party financing or credit support to acquire chips and develop capacity, and if such financing or support becomes unavailable or more costly, contracted capacity may be delayed, repriced, reduced, or not delivered.” — Anthropic amended Form S-1

“…our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.”” — Anthropic amended Form S-1

OpenAI Brings In Cathedral for Its Next Security Sprint and Raises Its Top Bug Bounty to $172,000

OpenAI is turning its own models on itself again, with outside help this time. OpenAI’s post on preparing its next security sprint, dated November 2, says it is planning another internal sprint and is partnering with Cathedral, whose researchers will use OpenAI’s models alongside OpenAI teams to challenge its defenses. OpenAI says its last sprint dedicated 25% of its product engineers to finding and fixing weaknesses, using Codex to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review. It argues that traditional cyber defenses alone are no longer sufficient, and it repeats its commitment of $1 billion in subsidized access over six months for frontline defenders protecting essential services. The company did not give a start date or duration for the new sprint.

The outside-researcher side of the program got a large raise the same day. OpenAI’s announcement of bigger bug bounties says it is more than quadrupling the midpoint of its reward ranges for medium, high and critical findings on average. A critical vulnerability in ChatGPT, Codex or the API Platform now earns up to $30,000, and exceptional critical findings remain eligible for up to $100,000, rising to $172,000 with bonuses. New multipliers reward high-quality reports with a reliable proof of concept, a five-tier loyalty program rewards repeat researchers, and short priority campaigns add bonus pools for newly launched features. In one of OpenAI’s illustrative examples, a one-click ChatGPT account takeover that historically paid $6,500 would now pay $51,577. Unpaid reports get the new ranges, and anyone paid in the previous 30 days gets the difference. OpenAI’s security team will also be at the PROMPT2PWN event in Oakland on November 20 and 21.

“Every company should be thinking about how to put these capabilities to work.” — Gavin Kliger, CEO and co-founder, Cathedral

OpenAI Rebuilt Its Git Backend for Codex-Scale Work, Cutting Median Push Latency by Nearly 5x

Coding agents are breaking the plumbing underneath software teams, and OpenAI has shown what it did about it. OpenAI’s engineering post on scaling Git, by Aaron Friel, Taylor Blau and Ted Nyman, describes Pando, a new Git-compatible backend that now powers all development in OpenAI’s monorepo. The authors say Codex changed the scale of development, with more branches, pushes, merges and fetches, and that starting around the end of last year the old systems struggled, with reliability at times falling below 90%. In production, Pando cut median push latency by nearly 5x, or nearly 80%, and delivers nearly a petabyte of Git data per day. It has served more than 57 terabytes of Git packs in a single hour, with over 99% of those bytes coming from local disk cache.

The design gives up two things most Git servers treat as requirements. Pando coordinates each branch independently instead of ordering every reference update across the whole repository, and it reuses stored packfiles across fetches instead of building a custom one for each client, sending download links through Git’s packfile-uri feature. The New Stack’s analysis of the Pando backend explains that, by default, Pando advertises only the trunk branch as a base for feature-branch pushes, which makes each pushed pack reusable by other clients. OpenAI says Pando went from kick-off to production in 10 weeks, was written in Rust, was designed by a single engineer and implemented by teams of agents, and began serving production traffic in April. Agents using GPT-6 Astra wrote Lean and TLA+ specifications to find bugs and prove properties of the system. OpenAI has contributed several Git client fixes upstream and says it is exploring ways to share the technology with users.

“In practice, users and agents are agnostic to the global ordering of ref updates, and instead depend on the order of updates to a single reference.” — OpenAI engineering team

OpenAI’s LifeVal Benchmark Measures Whether ChatGPT Can Finish Everyday Chores, Not Just Advise on Them

OpenAI also published a different kind of yardstick on Monday. OpenAI’s LifeVal announcement describes an evaluation of everyday tasks such as finding childcare, accessing government services, creating a budget and ordering groceries. To pick them, OpenAI started from the American Time Use Survey and the Multinational Time Use Study, then ran a new demographically representative survey of 3,995 respondents about where people most wanted help from a personal assistant. The tasks cover caregiving, health, education, household administration, personal finances, shopping and government services, and OpenAI says it used the evaluation over the past year to track and improve its models on browser use, computer use and connectors. Example tasks on the page include enrolling a child in kindergarten after a move (93.1%), organizing flood-preparation volunteers (98.0%) and finding a home in Luzerne County, then adding the deed to a county records cart for review (100%).

AI Affairs’ summary of the LifeVal results reports that on OpenAI’s chart GPT-5.2 scored 49.3% average reward and GPT-6 Astra scored 80.6%. The benchmark is OpenAI’s own, and the LifeVal page does not publish the full task set or the grading rubrics, so outside researchers cannot yet reproduce the numbers.

“To be a truly helpful personal assistant, ChatGPT should be able to take actions that actually get things done.” — OpenAI

Anthropic Releases FlashProve, an Open-Source Prototype That Proves Which Model Wrote a Reply

One of the quieter releases of the day addresses a basic trust problem. Anthropic’s research post on provable inference for large language models, published November 2 with co-authors from EPFL, StarkWare, MIT and other institutions, introduces FlashProve, an open-source prototype that uses a zero-knowledge succinct proof, or zkSNARK, to let a provider commit to a model and then prove that a response came from it without revealing the weights. For Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model in 4-bit quantization, Anthropic says a single H100 GPU generates proofs at around 110 tokens per second, with a proof of about 19 MB per 1,024-token request that can be checked in seconds on a 24-core CPU. That is about 70 to 95 times the GPU time of running the model on vLLM, which Anthropic says is far faster than prior open-source systems. The post credits Claude with producing the improvements to zkSNARKs for inference.

The FlashProve source code repository on GitHub accompanies the post, which also ships a Lean formalization and a technical description. Anthropic is explicit that this is a research prototype. The proof covers an integer version of the weights under greedy decoding, so sampled or floating-point outputs can diverge; the system assumes a mixture-of-experts architecture; and Anthropic does not guarantee that the Lean protocol, the paper and the fast code all match. It also proves responses after the fact rather than in line with serving.

“A user querying a model today cannot check that the responses came from the specific model that a developer claims is responding, as opposed to one secretly substituted in by a routing service for cost reasons, or even by an attacker who has tampered with the token stream.” — Anthropic research team

Nvidia Reportedly Tells Cloud Partners to Sell at Least 51% of New Capacity to Non-Chinese Customers

The chip war moved into rental contracts. The Next Web’s report on Nvidia capping its partners’ Chinese business, citing the Financial Times and five people familiar with the matter, says Nvidia has instructed cloud providers to sell at least 51% of their services to non-Chinese companies. The rule applies to new contracts only, was communicated verbally through Nvidia’s sales force and regional teams rather than in writing, and is a restriction the company had not imposed before. Nvidia neither confirmed nor denied the move to AFP, and China’s foreign ministry said it was not aware of the report. Many Chinese technology companies rely on capacity rented from cloud providers outside China to train models, which Washington has tried to restrict.

Seoul Economic Daily’s account of the FT report adds that the requirement applies to customers globally, not just to data centers in Asia, and that a Singapore-based provider said it is hard to find enough non-Chinese customers outside the United States. It also says a rack of Nvidia B300 chips that sells for about $600,000 internationally traded in China last month at 18.5 million yuan, or about $2.8 million, and that Nvidia had already more than halved the number of Asian customers approved to buy its AI chips after drawing up a whitelist this year.

“America cannot win the AI race unless America’s AI infrastructure wins worldwide.” — Nvidia statement, via The Next Web

Mistral Publishes the Technical Report Behind Le Chonk

Mistral followed last week’s open-weight release with the paper. The Mistral Large 4 technical report on alphaXiv describes a 52-billion-active, 1-trillion-total-parameter multimodal mixture-of-experts model and an agent-based pipeline that writes and verifies executable training environments for software engineering, cybersecurity and knowledge work. According to the abstract, efficient serving combines sparse attention, optimized context parallelism, prefill–decode disaggregation and speculative decoding. Calibrated NVFP4 quantization reduces the checkpoint to 606 GB, enabling deployment on four GB300 GPUs, and distillation provides high, medium and low reasoning-effort modes. The Mistral AI organization page on Hugging Face lists the Mistral-Large-4-1T-A52B weights alongside an NVFP4 variant.

Anthropic Opens a Paid Residency for Senior Technical Leaders Moving Into AI Safety

Anthropic is also recruiting from outside machine learning. Anthropic’s Alignment Science post opening the Anthropic Residency invites professors in quantitative disciplines, technical founders and staff-level engineers to apply for paid, fixed-term placements of six weeks, three months or six months, with no machine learning background required. Residents are paired with an Anthropic mentor, receive a dedicated compute budget and are paid the equivalent of $350,000 to $500,000 a year. Anthropic prefers work from San Francisco but allows remote participation, especially for shorter placements. Applications for this round close on November 22, 2026.

What We Left Out

Posts on X claimed that DeepSeek plans to pre-train its next V5 model on Huawei Ascend chips, citing a STAR Market Daily report based on supply-chain sources. DeepSeek has not confirmed it, so we treat it as a rumor. Separate posts again pointed to a “GPT-6.2 Astra” model string and teasing posts from OpenAI staff; OpenAI has announced no such model, and we found no primary source. We also skipped several funding announcements and product demos that we could not confirm against a primary source by Tuesday morning.