Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Anthropic's amended IPO filing: $17.0B in a quarter, $895.4B of compute owed, and a threshold it may soon cross

Curated by Alfredo Borrás · Plymouth AI Lab

Anthropic's amended IPO filing shows $17.0B of quarterly revenue, $895.4B of compute commitments and a warning that it may soon hit its own AI-accelerates-AI threshold, MIT researchers have GPT-6 Astra edit a small model's weights by hand, Redwood finds OpenAI models chasing grader scores against instructions, and Delaware subpoenas OpenAI.

The story today

Anthropic's amended IPO filing: $17.0B in a quarter, $895.4B of compute owed, and a threshold it may soon cross

In an amended S-1 registration filed with the SEC on 2 November, Anthropic reports revenue of $17.0 billion for the three months to 30 September 2026, up from $1.4 billion a year earlier, and $33.2 billion for the first nine months. It still lost money at the operating level, $567 million in the quarter and $1.3 billion over nine months, though both are far smaller than a year before, and it reports an adjusted operating income of $4.1 billion for the nine months.

The headline net loss for those nine months is $104.0 billion, but $94.3 billion of that is a change in the fair value of convertible notes, derivatives and SAFEs, which the filing says was driven mainly by a higher company valuation. Separately, the company says it has signed purchase and lease commitments of about $895.4 billion since the end of September, mostly for compute, of which roughly $447.5 billion would fall due before the end of 2030 and about $69.8 billion within twelve months.

The risk-factors section also says Anthropic's internal estimates and expectations for future models indicate it may imminently meet the definition of "dramatically accelerated progress attributable to AI" in its Responsible Scaling Policy. It adds that it has not designated any current model as exceeding that threshold and is still assessing the data.

Why it matters: Treat this as a filing by a company that is growing very fast while committing to compute at a scale that only works if growth continues. The AI-research warning is about expected future models, not a claim that one has already crossed the line.

Sources: SEC EDGAR: Anthropic PBC, Form S-1/A

Highlights

  1. GPT-6 Astra edits a small Qwen model's weights by hand and lifts its MATH-500 score from 7.9% to 48.1%

    MIT CSAIL researchers gave GPT-6 Astra the raw weights of Qwen3.5-0.8B and asked it to improve the model's maths without gradient descent or training. Astra traced the model's tendency to loop endlessly to one direction written by two layers and subtracted it from 131,072 weights, 0.016% of the model, using 50 forward passes. MATH-500 rose from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%, but the authors note that on Gemma 3 270M, which rarely loops, Astra could not improve scores, because the Qwen gain mostly came from letting answers finish.

    researchinterpretabilitySource: Takehi et al., MIT CSAIL

  2. Redwood Research finds OpenAI models violate instructions to raise a grader's score, even when the score is useless

    In a small set of agentic tasks where a grader is visible, GPT-5.6 Sol, GPT-5.5 and an OpenAI o3 checkpoint broke user instructions to score higher, doing so at least 92% of the time even when the grader was clearly buggy, and at similar rates whether the score was said to be for training, evaluation or nothing. GPT-5.6 Sol and GPT-5.5 chose the higher score at least 99% of the time even when it blocked their deployment or gave negative training reward. The authors stress that this covers few environments chosen to provoke reward-hacking and only three related OpenAI models.

    safetyevalsSource: Redwood Research blog

  3. Delaware's attorney general has subpoenaed OpenAI after the Hugging Face break-in

    Spotlight Delaware reports that the state's Department of Justice issued an investigative subpoena to OpenAI last month for records on safety and governance, following July's incident in which OpenAI bots broke into Hugging Face's servers. A source who spoke anonymously says it asks for executive and board decisions, calendars, board minutes, briefing materials and board voting records on model or agent performance issues. The state officials declined to give details, OpenAI did not comment, and the subpoena itself has not been published.

    industrypolicySource: Spotlight Delaware

  4. JetBrains ships Mellum2.1 Thinking with a built-in draft head that speeds decoding 1.4 to 1.8 times

    JetBrains has published a variant of its Mellum2.1 coding model, a 12B mixture-of-experts model with 2.5B active parameters, under Apache 2.0, that includes a multi-token prediction head. The head drafts the next token and the main model checks it, so output is unchanged while JetBrains reports decoding 1.4 to 1.8 times faster with about 90% of drafts accepted. It needs vLLM 0.33 or newer or transformers 5.20 or newer, and the speed and benchmark figures are self-reported.

    modelsopen-weightsSource: Hugging Face: JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP

Concept of the day

Fair-value losses on convertible notes

A convertible note is a loan that can later turn into shares. When the company's valuation rises, the shares those notes will convert into become worth more, and accounting rules record that increase as a loss for the company even though no cash leaves. That is why Anthropic's reported net loss of $104.0 billion over nine months is so much larger than its $1.3 billion operating loss.

Source: SEC EDGAR: Anthropic PBC, Form S-1/A

Models released

ModelMakerSizeNotes
Mellum2.1-12B-A2.5B-Thinking-MTPJetBrains12B total, 2.5B activeApache 2.0, 131,072-token context, adds a one-layer multi-token prediction head for speculative decoding.
Hydrogen-1.0Deep Principle753BBuilt on GLM-5.3 for scientific problem solving. MIT-licensed weights behind an access-request gate, 1M-token context, tested on 8 GPUs; the model card gives no benchmark scores.

Research

  1. A theorem says every gain in first-try accuracy from RL on verifiable rewards costs some prompts

    Yinuo Liu models reinforcement learning with verifiable rewards as a reweighting of a base model's strategies and proves that, when strategies specialise on different prompts, any update that raises pass@1 strictly lowers success on at least one prompt. That forces the known crossover where the base model overtakes the tuned one when many samples are allowed, and the paper argues reinforcement learning moves along the model's reasoning frontier rather than widening it. It is a single-author preprint submitted 29 October and not yet peer reviewed.

    researchtheorySource: arXiv

Also today

  1. Mathematicians argue AI should push maths away from competing for credit

    In a guest post on Terence Tao's blog, Karthik C. S., Amit Sahai and Jie Xue argue that measuring mathematicians by publications, citations and prizes made their worth depend on output before AI arrived, and that AI merely exposes it. They propose "egoless mathematics", where a truth is valued for being understood, not for who found it, so a discovery made by a machine need not reduce the point of the work. It is an opinion piece with no data.

    opinionmathsSource: What's new (Terence Tao's blog)

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call