Frontier AI's costs went public this week, and its strongest models stayed gated

Between October 25 and October 31, three things happened at once. Frontier AI's costs showed up in securities filings and earnings releases. The most capable new models reached vetted groups before anyone else. And the models developers can buy today got cheaper and more open. These are not separate stories. The labs are spending more, gating their strongest systems more tightly, and competing harder on price in the tier below.

The bill for the frontier went public

Anthropic's S-1, filed October 26, gave the first audited view inside a frontier lab. Revenue for the first half of 2026 was $16.29 billion, up from $1.16 billion a year earlier. Compute and infrastructure cost $8.95 billion over the same six months. Anthropic has applied to list Class A shares on Nasdaq as ANT. A founders' LLC would hold 50.1% of total voting power, so outside shareholders would not control the company.

A day later, Meta's pricing term sheet raised its common-stock offering to $25.0 billion at $625 a share. It came alongside a mandatory convertible preferred offering expected to net about $12.5 billion. NVIDIA said it will put $3 billion into Meta's offering and expand their work together on the Muse models. On October 28, Microsoft reported fiscal first-quarter revenue of $92.3 billion, up 19%. Azure grew 48% to $33.2 billion, and additions to property and equipment were $39.4 billion in the quarter.

Together these filings make one point. Anthropic's spending on compute alone exceeded half its revenue. Meta is raising tens of billions of dollars in new equity, with NVIDIA among the buyers. Microsoft's quarterly spending on equipment now exceeds Azure's quarterly revenue. Whatever the models cost per token, they are being built on capital spending at a scale few companies can match.

The strongest models arrived behind gates

Google DeepMind announced Gemini 4 Carbon on October 28. Google says it ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. But Google is releasing it first to a subset of cyber defenders, with additional trusted partners getting access in phases. It gave no date for general developer access. The Decoder reports that the model is also going through the U.S. government's voluntary pre-release testing. On October 30, OpenAI's API changelog listed GPT-6 Rosalind, a model for biological research that only organizations approved through a trusted-access program can use. Genie 4, DeepMind's real-time world model, launched on October 27 as a research preview for trusted testers only.

Gated access is not new, but this week it was the default for frontier systems. Their capabilities are announced first, then made available to selected users, and only later released to everyone, if at all. Labs justify the gate on cyber and bio risk. That claim cannot be checked from outside, and neither can the benchmark lead Google reports for Carbon, which comes from Google's own announcement.

Below the frontier: cheaper and more open

The tier developers can actually use moved the other way. OpenAI released GPT-6.1 Terra on October 26 for computer use, visual reasoning and writing. Its documentation lists $1 per million input tokens, $0.05 cached input and $5 output. The context window is 1.05 million tokens, with at most 922,000 tokens of input and 128,000 of output. Tool calling requires the Responses API. On October 28 the changelog added that eligible ChatGPT Pro and Business subscribers can claim up to $500 a month in API credits by linking an API organization they own.

On October 29, Mistral published the weights of Mistral Large 4, three weeks after its public preview. The model card describes a Mixture-of-Experts model with about 1 trillion parameters and 49 billion active per token. The license is a modified MIT license. Anyone may use, modify and commercialize the model, except companies that sell it as a hosted service, which need a separate license. Mistral's hosted API charges $0.70 per million input tokens and $2.10 per million output tokens.

For model selection, the gap is widening. The models with the strongest vendor-reported results are the hardest to get. The models you can deploy this week are getting cheaper, and in Mistral's case you can run them on your own infrastructure. For most agent and coding workloads, Terra-class and open-weight models are worth testing on your own tasks before you wait for gated access.

Rules and housekeeping

OpenAI published a new version of its Model Spec, dated October 29. It is organized around five core values: being faithful to human intent, minimizing harm, being honest and transparent, facilitating human agency, and acting with care. For anyone writing system prompts for OpenAI models, the Spec sets the order in which conflicting instructions are resolved, so it is worth rereading. Separately, as scheduled, existing OpenAI evals became read-only on October 31. According to OpenAI's deprecations page, the Evals dashboard and API shut down on November 30, so teams still using them have one month to export.

One infrastructure milestone stood out. Google said its Project Suncatcher prototype satellite ran a prompt on its onboard Gemma 4 model and returned the output to the ground. This is an engineering test, not new capacity anyone can use.

What to take into next week

This week had no single headline release. What it did have was disclosure: of what frontier AI costs, and of how access to the most capable models is being rationed. The practical takeaway is to benchmark against what you can actually deploy. Treat leaderboard positions you cannot reproduce yourself as claims, not results.