GPT-6.1 Terra: a faster step in computer use, at half the price

On October 26, OpenAI released GPT-6.1 Terra: a fast, efficient model combining detailed visual reasoning, precise computer use and clearer writing, described as a significant step up from GPT-5.6 Terra. It targets the most ordinary weekday work — drafting proposals, interpreting technical drawings, filling forms, editing videos. The API name is gpt-6.1-terra at $1 per million input tokens, $0.05 cached and $5 output — standard prices 50% below GPT-5.6 Terra and GPT-6.1 Sol.

Inside ChatGPT Work and Codex, the same usage allowance covers more tasks than GPT-5.6 Terra — “substantially more than GPT-6.1 Sol or GPT-6 Astra” in the official wording — for Plus, Pro, Enterprise and Edu users. The same day, GitHub Copilot added it across its surfaces: VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse, rolling out gradually on Pro/Pro+/Max/Business/Enterprise, billed at provider list prices under usage-based billing, with admin model-policy control. Same-day adoption across a coding toolchain is among the fastest in the records on this site.

• Read the official announcement ↗
• See the Copilot access conditions and plans ↗

The exam the vendor set for Terra is concrete: Agents' Last Exam, a set of professional computer tasks — for example turning a set of architectural drawings into a 3D model, or reviewing a feature-length animated film shot by shot into a time-coded editorial log. At medium reasoning effort, Terra scores 15 percentage points over GPT-5.6 Terra at about a third of the cost; it matches GPT-6 Astra's score at roughly a tenth of the token price. Speed enters usable territory too: tasks complete about 3x faster than GPT-6 Astra and GPT-6.1 Sol.

At maximum effort the bill keeps falling: against Claude Opus 5.5, Terra is about 3x more token-efficient at a quarter of the price per token — roughly 9x lower cost per task. The vision and desktop references point the same way: MMMU-Pro visual reasoning beats GPT-5.6 Terra at every setting while nearly matching GPT-6 Astra and GPT-6.1 Sol at much lower cost, and on OSWorld 2.1 desktop-app multitask work, medium effort already surpasses GPT-5.6 Terra's best score for a fraction of the cost. The vendor also states that safety evaluations meet the Critical threshold for cybersecurity under its Preparedness Framework.

• Next: when Sol 6.1 nears Astra, which work merits a switch →
• Next: the Sol, Luna and Opus 5.5 cost comparison →

More practical for daily use is the error-rate set. On internal error-inducing prompts, GPT-5.6 Terra at low reasoning effort errs 53.6% of the time; GPT-6.1 Terra at the same setting drops to 16.7%, a 37-point gap. At maximum effort it reaches 4.3%, 18 points below its predecessor and, per the vendor, nearly matching GPT-6 Astra and GPT-6.1 Sol. The same tasks run cheaper with fewer errors — the part of Terra's positioning that most directly touches real workflows.

Coding and research terminals need a split reading. Terminal-Bench 4.0 at max effort reaches 50.3%, more than double GPT-5.6 Terra's 21.5%; yet the vendor explicitly recommends GPT-6.1 Sol and GPT-6 Astra for the most complex coding. On Terminal-Bench-Science 0.1, max effort passes GPT-6.1 Sol at lower cost and lands within 5 points of Claude Opus 5.5 at less than a fifth of its cost. In other words: budget-sensitive, well-scoped agent work goes to Terra, while the heaviest coding and reasoning stay with Sol, Astra or Opus — the same task-by-tier thinking this site has recorded before.

• Next: how one browser task's cost was checked →

To use it today there are three routes: ChatGPT Work and Codex (Plus/Pro/Enterprise/Edu), the gpt-6.1-terra API, and GitHub Copilot's surfaces (gradual rollout — if the model is missing, wait a few days). Copilot bills usage at provider list prices; inside ChatGPT Work and Codex it draws on plan allowances rather than per-token fees.

On the independent side the gap is stated plainly: our October 27 reread of the full Artificial Analysis intelligence board — all 31 catalogued rows, matching the October 26 observation — still shows no gpt-6.1-terra entry; the older GPT-5.6 Terra (max) sits at 42 there. Terra is less than a day old, so the lag is expected — the +15 points and 50.3% above are vendor figures. When AA publishes a same-method score, this site will add it under the board's rules rather than pre-computing a rank. On-site references: GPT-6.1 Sol's AA 52 and the switching advice, the Sol/Luna/Opus cost comparison, and the Astra topic's browser-task cost check all pair well with this piece.

• See the model catalog and three independent boards →
• Check the AA leaderboard source ↗