OpenAI releases GPT-6.1 Terra model for computer control

According to OpenAI, in an internal test on error-inducing prompts, the share of incorrect answers at a low reasoning effort dropped to 16.7% compared to 53.6% for GPT-5.6 Terra. At the maximum level, the figure was 4.3%: the company considers it comparable to the flagship GPT-6 Astra and GPT-6.1 Sol models. In the Agents' Last Exam benchmark, where AI agents solve professional tasks in software, Terra scored 15 percentage points higher than its predecessor at medium reasoning effort, and 60% at maximum effort with a cost of $1.01 (compared to 50.7% for $3.58 for GPT-5.6 Terra, and 63.2% for $8.82 for Anthropic's Claude Opus 5.5). In Terminal-Bench 4.0, the score increased from 21.5% to 50.3%.

API pricing is $1 per million input tokens, $5 per million output tokens, and $0.05 per million cached input tokens. According to the company, the price has been halved compared to GPT-5.6 Terra and GPT-6.1 Sol. The model's context window is 1.05 million tokens, with a maximum output size of 128,000 tokens.

Plus, Pro, Enterprise, and Edu subscribers can access Terra in ChatGPT Work and Codex, while developers can use it via the API under the identifier gpt-6.1-terra; the model will not be available in standard ChatGPT chat. For complex coding tasks, OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra. In its own risk assessment framework, the company classified Terra as having a critical cybersecurity risk level and maintained the safety mitigations developed for GPT-6 Astra. All benchmark results are reported by OpenAI itself; as Kod.ru notes, independent evaluations are not yet available.