OpenAI and Google launch new AI models with opposite rules on who can use them

Official statements. On 26 October OpenAI released GPT-6.1 Terra in ChatGPT Work, Codex and the API, positioning it as a fast, cheaper model for computer use, visual reasoning and writing. OpenAI says its API prices are half those of GPT-5.6 Terra, that it solves 50.3% of Terminal-Bench 4.0 at maximum effort, and that it meets the “Critical” cybersecurity threshold of its Preparedness Framework. All figures are OpenAI’s own. [1]

Two days later Google DeepMind announced Gemini 4 Carbon, two weeks after Gemini 4 Argon. Google claims the top score on the Artificial Analysis Intelligence Index v5.0 and leads on vulnerability benchmarks (SEC-bench Pro 88.2%, ExploitGym 54.8%). It is releasing Carbon first to a subset of cyber defenders in its Fairwind Program and takes part in the US government’s voluntary pre-release testing. Broader availability has no date. [2][3]

Signals on X. Sam Altman framed Terra as the start of “this new era of computer use”; Demis Hassabis wrote that Google is “iterating at lightning speed” and that Carbon is “SOTA across key benchmarks”. Both are first-party claims and neither post names independent evaluations. [4][5]

Independent checks. SecurityWeek notes that Google’s cyber results were obtained with no guardrails in place. [3] On the maintainer-run Terminal-Bench board we track, Terra’s self-reported 50.3% would sit below the top verified agent entries. Our reading: the labs are now competing on deployment policy as much as on scores — one widens access by cutting price, the other narrows it because of what the model can do.

Signal concentration: both X signals and both companies are US/UK. Non-US counterparts this week included open agent research from China, covered in Agents.