AivexaNewsSearch
AI news for builders and product teamsChecked every hour

AI #192: Anthropic Is Going Public

Collected Oct 29, 2026

Anthropic filed an S-1 this week to prepare for an initial public offering. The filing leaves many numbers to be filled in later, including the share price, but it already exposes how the company's messaging to investors diverges from its other public rhetoric and how its corporate control is structured. This is the first formal step toward a public listing for one of the leading frontier AI labs, and it lands in a period when Anthropic is also expanding its compute commitments.

On the money side, Anthropic has secured a new $63 billion debt deal tied to TPUs, and its growth reportedly picked up again after the release of Opus 5.5. The combination is notable: a company heading toward public markets is pairing rapid revenue momentum with very large fixed obligations for custom silicon. In practice, a filing gives outside investors their first structured view of that balance sheet, though the S-1 as described still has gaps on valuation.

Google announced Gemini 4 Carbon, which tops the new Artificial Analysis Intelligence Index v5. Carbon arrives roughly two weeks after Gemini 4 Argon, which the author had just finished reviewing. The catch is availability: the announcement came without general access, with entry of queries reportedly weeks away. Argon itself was assessed as a strong but not fully frontier model, stronger than any non-Anthropic, non-OpenAI model with the possible exception of Kimi K3.1 Pro, with a low hallucination rate, top placement on Scale Labs' HSS benchmark and on Vals and text arena, and a Minecraft run that defeated the Ender Dragon in 27 hours after writing 2,783 programs across four agents. Complaints centered on low usage limits and small effort and context settings.

Artificial Analysis moved to Intelligence Index v5, its biggest upgrade since v4 launched in January, with 60% of the index now built on private test sets or held-out answers. Two benchmarks are new: FrontierCode 2.0 from Cognition, which grades agent coding changes on whether a repository maintainer would merge them and uses deliberately underspecified prompts, and Terminal-Bench-Science, 70 agentic research tasks run in a terminal. HLE-Diamond replaces Humanity's Last Exam as a refined 1,000-question set split evenly between reasoning and knowledge, AA-Omniscience v1.1 removes saturated and public questions, and CritPt v2 is a private 40-question hard subset run with code execution. GDPval-AA v2.1, SciCode and AA-LCR v1.1 were removed. A companion Progress Index v1 tracks capability over time on one continuous scale.

The v5 results put Claude Opus 5.5 (max) first at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55), with Gemini 4 Argon (high) next at 50 and Anthropic and OpenAI holding the top five. The open-weights gap widened from 3 points to 12: Kimi K3.1 Pro, whose weights are not yet released, leads that group at 46, while among released weights DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32, below Claude Haiku 5.5 (max, 36). Cost per task varies more than 13-fold among the top ten, from $0.96 for GPT-6.1 Terra (max) to $13.04 for Claude Fable 5.1 (max).

Those numbers drew objections. Critics argued the index penalizes open models by construction, and one complaint held that FrontierCode 2.0 uses different harnesses for commercial and open models, muddying whether model or harness is being measured. The author's reading is that Chinese labs have been targeting benchmarks rather than that the index is rigged. A sensible portfolio from the new ratings, per the author: Opus 5.5 where quality dominates, GPT-6.1 Sol for easier tasks, GPT-6 Astra for narrow strengths, and DeepSeek V4.1 Flash 1017 or MiMo-v2.6-Pro as cheap open alternatives.

Other releases include GPT-6.1 Terra, roughly on par with 6.1 Sol after adjusting for model size and recommended by Ted Sanders for communication, vision and computer use, and Kimi K3.1 Pro at $2/$10 per million tokens with $0.20 cache, or $4/$20 for high speed. Moonshot said it is intentionally throttling cyber capability, keeping K3.1 Pro's offensive cyber ability below its pipeline's ceiling and behind frontier proprietary models, while planning to raise open-weight cyber capability over time. DeepMind's Genie 4 also shipped, to positive reception.

OpenAI announced three research results: local smoothing and the Kakeya conjecture in all dimensions, described as plausibly Fields Medal material; a disproof of the Exponential Time Hypothesis, long widely believed true, with verification shepherded by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT, raw data published and a conference funded; and a rigorous existence proof of interacting quantum field theories with gauge fields and fermions in four dimensions, including QCD and a unified completion of the Standard Model. Anthropic posted its own ETH refutation found purely via Claude, dated the same October 16, at roughly exp(O(n/√log n)) like OpenAI's. An estimated 800-plus papers relied on ETH; OpenAI said cryptosystems are unaffected for now.

Security dominated the rest of the week. Red teamers reached Anthropic's Slack through third-party vulnerabilities, the scariest incident so far; both OpenAI and Anthropic build heavily on Slack. Most important hacks came from red teamers and white hats, with no major damage yet. GMO refused a large ransom demand. METR red teamed OpenAI's internal agent monitoring, with more work left. A flaw in the OpenAI API used for fraudulent distillation was fixed. Separately, Microsoft offered buyouts to 60,000 workers, and a lawsuit targets OpenAI over GPT-4o helping a sick user taper medication while advising her to conceal it from doctors.

Why it matters: Developers get a new leaderboard that weighs private test sets heavily and a clearer view that frontier capability is concentrated in two labs, with the open-weights gap at 12 points. Commercial buyers face a wide cost spread per task and should treat announcements like Carbon cautiously until access exists. Product teams planning on Anthropic now also have to consider the disclosure and governance changes a public listing implies, though what the S-1 ultimately reveals is, by inference, still unknown.

Read at Zvi Mowshowitz

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Anthropic filed its S-1 this week to prepare for its IPO.