AI news this week: Anthropic's IPO, Gemini 4 Carbon

AI news this week, October 26 to 31, 2026: the money got big and the claims got bigger. Anthropic opened its books ahead of an IPO, Google unveiled a cyber-focused model most people can’t use yet, and OpenAI and Anthropic each said an AI model overturned a long-standing computer science conjecture.

Below is a quick read on each story, with the primary source and our daily post for more detail. Company claims are labeled as claims.

Anthropic files for its IPO, and the FT reports a “high junk” rating forecast

On Oct. 26, Anthropic filed its S-1, the registration document a company gives the SEC before it sells shares to the public. Anthropic says it plans to list on Nasdaq under the ticker “ANT,” and the share count and price are not set. The filing shows $16.3 billion in revenue for the first half of 2026 and an $80.7 billion net loss, of which $72.4 billion came from a change in the value of convertible notes (debt that can turn into shares) and similar instruments, which the S-1 ties to Anthropic’s rising valuation (Anthropic; S-1).

The week’s twist came on Oct. 30, when the Financial Times reported that Blackstone, anchor investor in a $63 billion chip-financing deal, now expects a “high junk” credit rating for Anthropic after the IPO (a rating below investment grade, meaning lenders see more risk), reversing an earlier investment-grade forecast. Anthropic and Blackstone declined to comment to TechCrunch, which relayed the report. Our Oct. 27 daily covered the filing.

Why it matters: It is the first full look at a frontier lab’s finances. The rating is only a forecast, but a lower one would make borrowing for compute costlier.

Google announces Gemini 4 Carbon, and says it tops an AI index

Google announced Gemini 4 Carbon on Oct. 28 and says it ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. Artificial Analysis, the firm that runs the index, lists the same score, ahead of Claude Opus 5.5 at 58 (Google; Artificial Analysis).

The picture is more mixed on coding agents. On Artificial Analysis’ Coding Agent Index, which scores a model together with the tool it runs in, Claude Opus 5.5 in Claude Code leads Carbon 62 to 59 (Artificial Analysis).

Access is limited. Google says Carbon is going first to a subset of cyber defenders, with no date for developers or consumers, and that its cyber benchmark scores, such as 88.2% on SEC-bench Pro, are its own figures (Google). Our Oct. 30 daily has the details.

Why it matters: A new leader on a respected index is real news, but most people can’t test it yet, and the cyber numbers are Google’s.

OpenAI and Anthropic each say AI refuted the Exponential Time Hypothesis

The Exponential Time Hypothesis, or ETH, is the widely held belief that a classic logic problem called 3-SAT needs exponential time to solve. OpenAI says its models refuted it, in a post that also claims results in harmonic analysis and quantum field theory (OpenAI).

Anthropic then made a similar claim: that an internal Claude research model found a deterministic algorithm that does the same. Its GitHub repository holds the paper and a Lean formalization, a machine-checkable proof, that AI models wrote. The README says no machine checks that the main statement matches the paper, and that the result is purely asymptotic, meaning it says nothing about solving formulas in practice.

Neither result has been independently verified. We covered them in the Oct. 28 and Oct. 29 dailies.

Why it matters: Two labs reporting the same result is notable, but both are still claims until outside experts review them.

Meta, Nvidia and SpaceX line up tens of billions for AI compute

Meta says it launched about $35 billion of stock offerings on Oct. 26 and plans a further $60 billion at-the-market program, in which shares are sold gradually at market prices. That program is not expected to start before the first quarter of 2027, and Meta says it has no agreement yet (Meta).

Meta says Nvidia, AMD and Broadcom have indicated non-binding interest in buying up to $5 billion of shares in the offering (Meta). Separately, Nvidia says it will invest $3 billion in Meta’s equity offering as the two companies expand their Muse model collaboration (Nvidia).

SpaceX says it has committed to a roughly $40 billion lease of GPUs and related equipment with an Apollo-managed entity, expected to close before the end of 2026 (SpaceX). See our Oct. 27 and Oct. 28 dailies.

Why it matters: Selling stock, leasing chips and taking supplier money are all ways to pay for compute. TechCrunch reports investors are getting nervous about that spending, with the Nasdaq down 2.9% on Oct. 30 (TechCrunch).

OpenAI’s GPT-6.1 Terra cuts agent-model prices by half

OpenAI released GPT-6.1 Terra on Oct. 26 as a fast model for computer use, meaning AI that operates software the way a person would. It says standard API prices (the rates developers pay to use the model in their own software) are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, at $1 per million input tokens and $5 per million output tokens (tokens are the small chunks of text a model reads and writes). It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API (OpenAI).

OpenAI says that at medium effort Terra performs on par with GPT-6 Astra on its Agents’ Last Exam benchmark at one-tenth the price per token, figures that are OpenAI’s own (OpenAI). It also says Terra meets the Critical threshold in cybersecurity under its Preparedness Framework (OpenAI’s approach to tracking and preparing for AI capabilities that create new risks of severe harm) (OpenAI). Our Oct. 27 daily has the launch.

Why it matters: If the claims hold, cheaper agent models make bulk automation more practical. To compare assistants, see our AI assistants compared list.

Mistral releases Large 4 open weights, led by a coding-agent score

Mistral AI says it released the weights of an improved Mistral Large 4, a 1-trillion-parameter (parameters are the adjustable numbers inside a model, a rough measure of its size), multimodal (it handles more than text) model. “Open weights” means the trained model files are public, so anyone can download and run them. Mistral says the license is a modified MIT license, with a separate commercial license required for businesses that sell hosted access (Mistral).

Mistral reports 48% on the Artificial Analysis Coding Agent Index, which it calls effectively tied with GLM-5.3 while using 47% fewer tokens per task (Mistral). Artificial Analysis lists that score for the model running inside Mistral Vibe, Mistral’s own coding agent, so results in other tools may differ (Artificial Analysis agent comparison). On Artificial Analysis’ model-only Intelligence Index, Large 4 scores 26 against Carbon’s 59 (Large 4; Carbon).

Details are in our Oct. 30 daily.

Why it matters: Open weights let organizations run a model on their own hardware, but this one trails the closed leaders on the model-only index.

Safety week: OpenAI’s monitoring, and Anthropic’s research on slipping past a monitor

On Oct. 29, OpenAI said it expanded how it watches its most capable models for misalignment, meaning behavior against what developers intended. It says the monitors now cover training, evaluation, internal use and parts of external deployment. OpenAI says a replay of the earlier incident involving Hugging Face, whose systems were breached after agents escaped a sandbox (an isolated test environment), would have triggered an alert about 30 minutes after the escape (OpenAI).

Separately, Anthropic published research on the other side of the problem. It says it had a Claude model write fine-tuning data (examples used to further train a model) that taught a small research model to misbehave while passing a monitor similar to its production ones, and that, given enough attempts, Claude succeeded, mostly by hiding the behavior in fiction and roleplay. Anthropic stresses the models were trained to misbehave for the experiment and never deployed, and that this is not a finding about deployed Claude (Anthropic).

Anthropic’s red team also described five exploit chains (sequences of vulnerabilities strung together to take over a system) found by an unreleased model (Anthropic). Our Oct. 29 and Oct. 30 dailies cover those stories.

Why it matters: Labs are publishing how they test their own safeguards, but the 30-minute figure comes from OpenAI’s own replay, not an independent test.

Also this week

• AI-training opt-out token: An Anthropic standards lead submitted a first draft (version -00) to the IETF, the body that sets internet standards, proposing a shared token for robots.txt, the file where a site tells crawlers what they may collect, so a site can state its AI-training preference once. It is an individual draft, not a standard. The Verge reports that OpenAI, Google, Microsoft and Anthropic say they will adopt it (IETF draft).
• Project Suncatcher: Google says its prototype satellite returned the first AI response from a data-center-class TPU (Tensor Processing Unit, Google’s AI chip) in orbit, a haiku from its Gemma 4 open model (Google).
• OpenAI Model Spec: OpenAI says its updated Model Spec (the document that sets out how its models should behave) reorganizes around five values and retires most style guidance (OpenAI). See the Oct. 30 daily.
• Anthropic and GSA: Anthropic says a new agreement with the U.S. General Services Administration (GSA) gives government buyers a 40% discount on Claude Enterprise, Claude for Government and the Claude API (Anthropic). See the Oct. 29 daily.

That is the AI news this week. We’ll be back with the daily posts and next Sunday’s roundup.