OpenAI's GPT-6.1 Terra gains speed and lowers price

OpenAI updates Terra with GPT-6.1, an intermediate model focused on screen analysis and task execution. It becomes cheaper, faster, and remains reserved for certain ChatGPT subscribers.

OpenAI has introduced GPT-6.1 Terra, an intermediate model designed to read what appears on the screen and then act within software. Available since October 26, it targets work uses where AI must click, fill in, extract, or write without delay. The novelty is mainly notable because it combines improved performance, lower price, and already open access for the relevant subscribers.

Terra passe du simple copilote à l’exécutant appliqué

GPT-6.1 Terra succeeds GPT-5.6 Terra and gains a very concrete function: understanding a computer interface to chain actions. OpenAI describes it as a model capable of drafting an email, interpreting a technical plan, extracting data from an invoice, entering information into accounting software, or assisting with video editing. In short, it does not just respond; it acts. And it does so with what the company presents as better writing, more natural and clearer.

This evolution fits within OpenAI's range, which also includes Luna, Sol, and Astra. Terra occupies an intermediate position: more affordable than high-end models, but more robust than entry-level ones. OpenAI emphasizes the blend of detailed visual reasoning and precise computer use, a duo that interests product teams as much as admin, support, or creative roles.

Des gains mesurés sur les benchmarks, avec un coût qui descend

On Agents’ Last Exam, OpenAI says Terra outperforms GPT-5.6 Terra by 15 points in average reasoning mode. The company adds that at maximum effort, Terra is over three times more token-efficient than Claude Opus 5.5, with a cost per task about nine times lower [to be verified on exact methodological details]. On the same ground, OpenAI also claims Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

Pricing follows the same logic. GPT-6.1 Terra is priced at 1 dollar per million input tokens and 5 dollars per million output tokens, half the price of GPT-5.6 Terra according to OpenAI. Reading this pricing grid, Terra occupies a rather rare zone: performant enough for serious work, but without the entry ticket of the most powerful models. It is not the first “reasonable” model on the market, but OpenAI clearly wants to give it the best performance-price ratio possible.

Pour les entreprises, le vrai sujet reste le contrôle

The model is already available in ChatGPT Work and Codex, but only for ChatGPT Plus, Pro, Enterprise, and Edu subscribers. OpenAI thus reserves it for professional and semi-professional uses, where automation directly impacts screen time. The issue is not trivial: when a model can manipulate an interface, errors, blockages, and unexpected behaviors must also be monitored.

OpenAI indicates that Terra reaches the Critical threshold of its Preparedness framework for cybersecurity and retains the protections associated with GPT-6 Astra. The company also says it has observed better compliance with instructions, more transparency about the model’s limits, and improved resistance to prompt injection attempts. On this front, the AI remains a very fast intern: useful, but still capable of forgetting instructions when the context becomes tricky.

The release is finally accompanied by a system sheet, a sign that OpenAI wants to document the model’s behavior more precisely. For teams evaluating the adoption of agency tools, this type of publication matters as much as the raw score: it allows judging reliability, safeguards, and the real maturity level before wider deployment.

By the numbers

  • 15 points — gain on Agents’ Last Exam in average mode, according to OpenAI.
  • 1 dollar — price per million input tokens, according to OpenAI’s pricing grid.
  • 5 dollars — price per million output tokens, according to OpenAI’s pricing grid.
  • 3 times — speed announced compared to GPT-6 Astra and GPT-6.1 Sol.
  • 4 offers — access reserved for Plus, Pro, Enterprise, and Edu.

To read