The Week the Frontier Got Cheaper, Gated, and Filed for an IPO

If you check AI news once a week, this was the week to check. The story was not a single product; it was three pressures moving at once. Prices fell, the most powerful models got harder to reach, and the company at the center of it all filed to go public.

OpenAI shipped GPT-6.1 Terra, a computer-use model priced around one dollar per million input tokens and five per million output, roughly half of what came before, with speed the company says finishes office tasks three times faster. When the cost of an agent step drops that far, automations that were too expensive to run start to pencil out. Then came the leak: GPT-6.2 Astra, which skipped the 6.1 name after an internal safety test reportedly went wrong, appeared in Codex greyscale and was already eating a fifth of real traffic by early November.

Google answered with Gemini 4 Carbon, which topped an independent intelligence index at 59 and leads on a security benchmark that found 88.2 percent of vulnerabilities in one test. It is being rationed through a vetted-access program rather than opened to the public API, the pattern every frontier lab now follows: hand the powerful model to a few trusted groups first. Anthropic, meanwhile, exposed Claude Mythos 5.5 and Fable 5.5 in source code, a specialized near-unguarded model and a million-token general flagship, while Mistral released the open-weight Le Chonk, Mistral Large 4.

Then the money map. Anthropic filed its S-1, disclosing 16.3 billion dollars in first-half revenue, up roughly 13x, at a reported 1.4 to 1.5 trillion dollar valuation. And in China, MOFCOM published its first export controls on frontier models and training data on November 3, a third regime beside the EU and US rules, which means any company running agents across borders now answers which model ran, what data moved, and whether anything crossed a restricted jurisdiction. The week's real headline was that capability, cost, and control are now moving on the same track.