Anthropic files to list on Nasdaq, OpenAI halves computer-use prices, and its models publish proofs of three open problems

Six items from the 24 hours to October 27, 2026. Anthropic put its finances on the public record ahead of a listing, OpenAI shipped a cheaper model for driving a computer, and the same company said its models had produced proofs of three open problems.

Source scope: the fixed set of public AI and tech accounts this channel monitors while the personal X connection stays unlinked. One of them, Benedict Evans, keeps its account protected, so nothing from it appears here.

Anthropic files to list on Nasdaq under the symbol ANT Anthropic's announcement, its Form S-1 on SEC EDGAR, and Anthropic's post · 17,093 likes and 7.17 million views at capture · October 26

Anthropic filed a registration statement on Form S-1 with the SEC on October 26, the step before an initial public offering 1. The company intends to list Class A common stock on the Nasdaq Global Select Market under the symbol ANT. The number of shares and the price range are not yet set.

The public filing follows six confidential submissions that began on June 1 2. The prospectus is a draft marked "Subject To Completion", and it carries the numbers the announcement leaves out.

Revenue was $41 million in 2023, $386 million in 2024 and $4.59 billion in 2025. In the six months to June 30 it reached $16.29 billion, against $1.16 billion in the same half of last year 2.

Operating expenses rose more slowly than revenue. The company lost $693 million from operations in the first half of 2026, against $2.13 billion a year earlier, on operating expenses of $16.98 billion. Compute and infrastructure took $8.95 billion of that and personnel $3.15 billion, including $2.2 billion of stock compensation 2.

The net loss for the same six months was $80.75 billion. Of that, $72.39 billion came from a change in the fair value of convertible notes, derivative liabilities and SAFEs, driven by the rise in Anthropic's valuation 2. The charge is non-cash, and it tracks what those instruments are now worth rather than money paid out. As of September 30 the company counted more than 11,900 customers spending at least $100,000 a year, more than 1,900 spending at least $1 million, and about 250 spending at least $10 million 2.

OpenAI's GPT-6.1 Terra matches its frontier model on computer use at a tenth of the price OpenAI's launch page and its four-post launch thread, with follow-ups from Sam Altman and Greg Brockman · 24,907 likes and 4.09 million views on the lead post at capture · October 26

GPT-6.1 Terra is OpenAI's new mid-tier model, built for computer use: reading what is on a screen and acting across applications, alongside visual reasoning and writing 3. It is available today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers as gpt-6.1-terra.

The benchmark OpenAI leans on is Agents' Last Exam, which sets professional tasks such as turning architectural drawings into a 3D model. At medium reasoning effort, Terra scores 15 percentage points above GPT-5.6 Terra at roughly a third of the cost, and OpenAI says it matches GPT-6 Astra, its frontier model, at a tenth of Astra's price per token 3.

Standard API pricing is $1 per million input tokens and $5 per million output, half what GPT-5.6 Terra and GPT-6.1 Sol cost 3. At maximum effort, OpenAI estimates a cost per task about nine times lower than Claude Opus 5.5.

Two other numbers matter for work that gets checked. On OpenAI's internal set of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% to 16.7%. On Terminal-Bench 4.0, which scores command-line coding agents, Terra solves 50.3% of tasks against 21.5% published for GPT-5.6 Terra 3. Every benchmark figure here comes from OpenAI, measured on benchmarks the company designed.

OpenAI places Terra at the Critical threshold for cybersecurity under its Preparedness Framework, with the safeguards built for GPT-6 Astra. Greg Brockman, the company's president, wrote that computer use with Terra "has crossed a qualitative threshold where it feels faster at using a computer than what i can do unassisted" 4.

OpenAI says its models proved three open problems, and let outside mathematicians set the terms of release OpenAI's research page, the released papers and Lean formalizations in the openai/math repository, and Greg Brockman's posts · 11,970 likes and 1.70 million views on Brockman's post at capture · October 26

On October 26 OpenAI published three results, one each in harmonic analysis, complexity theory and quantum field theory, all produced by its own models 5. For each, an outside expert chose how the result would reach the public: release the output, form a group to explain it, or defer to that group on everything.

The mathematics result settles Sogge's 1991 local smoothing conjecture in every dimension. Guth, Wang and Zhang had proved it for two dimensions in 2019; the models' proof covers all of them, and because the conjecture implies the Kakeya conjecture and Stein's restriction conjecture, both fall with it 5. The models produced the proof on October 18, and it was formalized in Lean, a proof assistant that machine-checks every step, the next day. The write-up and the formalization are both in a public repository.

The computer science result refutes the Exponential Time Hypothesis, which says that deciding Boolean satisfiability takes exponential time in the worst case. Hundreds of papers rely on the hypothesis to prove that other problems are hard 5. The Lean statement released with the paper gives a deterministic machine that counts the satisfying assignments of a k-CNF formula in subexponential time and polynomial space 6.

The physics result is an existence proof for certain interacting quantum field theories containing gauge fields and fermions in four dimensions, including quantum chromodynamics and a unified completion of the Standard Model. As a subproblem, it proves the existence of pure Yang–Mills theory, one of the Clay Mathematics Institute's Millennium Prize problems 5.

What has been published differs by field. The math output went out whole, the complexity paper and its Lean proof are public, and David Gross's group asked OpenAI to withhold the physics artifacts while it writes them up 5. Russell Impagliazzo and Ramamohan Paturi, two of the three people who proposed the Exponential Time Hypothesis, will verify the complexity claim with Ryan Williams, and OpenAI agreed to fund a conference on it. Lean checks that a proof follows from its assumptions, which is less than checking that the assumptions say what was meant. All three went out as preprints and repository files rather than through journal review. Brockman's summary of the complexity claim was "p still probably doesn't equal np" 7.

Mollick had GPT-6 Pro extend a published finding about video games and crime Ethan Mollick's post and its two screenshots · 365 likes and 26,668 views at capture · October 26

Ethan Mollick, a Wharton professor who studies how people use AI, describes a finding he likes: violent video games reduce crime, because the people who would otherwise be out are at home playing them 8. He says he had GPT-6 Pro revisit it.

By his account, the model downloaded fresh crime data and extended the earlier work. He reports that the 2013 launch of Grand Theft Auto V lined up with a dip in violence internationally, and that other games were more ambiguous 8.

What he posted is two screenshots and a paragraph, so the specification, the sample and the controls stay out of reach. A launch date that overlaps a fall in violence is an association, and separating one game from everything else that changed that year is the work a paper would have to do. The workflow is the part worth noting: a professor ran an extension of a published result through a general-purpose model and published the output himself.

The thinking you can read is a summary, and Mollick says the summary has grown too thin to supervise Ethan Mollick's two posts, Anthropic's distillation research and its API documentation · 335 likes and 19,233 views on the first post at capture · October 27

Mollick argues that the end of detailed thinking summaries in Claude and ChatGPT has made it harder to be a human in the loop 9. Weighing in on work in progress needs something to read, so hiding the trace pushes users toward automation rather than augmentation.

The mechanism is documented on Anthropic's side. Its API returns thinking as a summary, and the documentation's own example prints a block labeled "Thinking summary" 10.

Anthropic has said why. Summarizing Claude's reasoning before showing it to users was an early measure against distillation, because a full transcript teaches a smaller model more than a summary does 11. A model trained on full Claude Opus 4.8 transcripts scored 80.4% on SWE-bench Verified; one trained on summaries scored 71.8%.

The trade-off lands on whoever supervises the work. What is available to inspect is a condensed account of the reasoning, and the reader has to trust the condensation 10. Mollick's follow-up calls the hiding itself a failure of imagination: the traces had grown too long and too proprietary, and concealing them so the AI "just does 'magic'" is worth pushing back on 12.

Graham: not reading or writing will soon be as common as smoking Paul Graham's post · 10,007 likes and 623,194 views at capture · October 26

"We're about to enter a period where not reading or writing becomes as ubiquitous as smoking once was" 13. Graham, who founded Y Combinator, adds that people may come to see the cost as they did with smoking, and that this is far from certain because the cost may not be obvious.

The comparison carries one idea: smoking became common years before its harm became undeniable, and Graham expects reading and writing to follow the same lag. His post names the analogy and leaves the cost itself unstated.

It drew 10,007 likes and 623,194 views, the largest response to any of Graham's own posts in the window; the next, on UK university spin-outs, drew 1,204.