OpenAI Releases GPT-6.1 Terra: Half the Price, Twice the Accuracy in Code

OpenAI has introduced GPT-6.1 Terra, a mid-tier model that comes close to the flagships on agentic tasks at a much lower price. Standard API pricing is $1 per million input tokens, $0.05 per million cached tokens and $5 per million output tokens. That is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. On Terminal-Bench 4.0 at maximum effort, Terra solves 50.3% of tasks, while the previous version solved 21.5%. The model is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and to developers through the API as gpt-6.1-terra.

Computer instead of chat: what Terra is built for

OpenAI positions the model as a workhorse for office tasks: drafting commercial proposals, reading technical drawings, filling out forms, editing video. For developers, the example in the announcement is as practical as it gets: read incoming invoices and enter the details into an accounting system. For a job like that, peak intelligence matters less than the cost per thousand documents and speed.

On Agents' Last Exam, a benchmark of professional tasks, Terra handles converting architectural drawings into a 3D model and marking up an animated film frame by frame for a restoration team. At medium reasoning effort, the model scores 15 percentage points higher than GPT-5.6 Terra at roughly a third of the cost. Its results match the flagship GPT-6 Astra, but its price per token is ten times lower. At maximum effort, Terra uses more than three times fewer tokens than Claude Opus 5.5, and its tokens are four times cheaper. As a result, a task costs roughly nine times less. Terra also completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

Benchmark and safety numbers

• Agents' Last Exam (medium effort): +15 percentage points over GPT-5.6 Terra at ~1/3 of the cost; matches GPT-6 Astra at 1/10 of the price per token.
• Agents' Last Exam (maximum effort): needs 3+ times fewer tokens than Claude Opus 5.5, with a token price 4 times lower and a task cost ~9 times lower.
• Speed: tasks complete more than 3 times faster than on GPT-6 Astra and GPT-6.1 Sol.
• MMMU-Pro (visual reasoning): higher than GPT-5.6 Terra at every setting, nearly level with Astra and Sol.
• OSWorld 2.1 (multistep work in desktop applications): at medium effort, beats the best GPT-5.6 Terra result at a small fraction of the task cost.
• Errors on provocative prompts (internal test): at low effort 53.6% → 16.7%; at maximum effort, down 18 percentage points to 4.3%, almost on par with Astra and Sol.
• Terminal-Bench 4.0: 50.3% versus 21.5%.
• Terminal-Bench-Science 0.1: at maximum effort, beats GPT-6.1 Sol at a lower cost and comes within five points of Opus 5.5 at an estimated price of less than one-fifth.
• Safety: Critical level for cybersecurity under the Preparedness Framework; the safeguards from GPT-6 Astra are retained; resistance to prompt injection is higher; and the model more often stays within the scope of the task and is more honest about its limitations.

The gap in hallucinations at low effort is especially telling: from 53.6% to 16.7%. Models usually make the most mistakes in the cheap mode, and that is exactly how they have to run in mass automation. OpenAI is also explicit about the limits. For the most demanding programming tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.

How to work out the economics in practice

With $1/$5 pricing and cached tokens at $0.05, Terra fits well into pipelines with repeated context. That means a long instruction, an accounting policy or a form's field schema, sent along with every document. A rough estimate: an invoice of 3000 input tokens, including the scan, plus a 300-token response costs about $0.0045. Ten thousand invoices come to around $45 without caching. Real numbers depend on image size and the chosen effort level, so measure them on your own data.

The practical sequence is as follows. Run a sample of tasks on Terra at low and medium effort and compare error rates and cost per task. Route hard cases, such as poorly structured documents or long chains of code edits, to Sol or Astra. For agents that read other people's emails and web pages, test resistance to prompt injection separately on your own scenarios. The results OpenAI reports come from internal and specialized benchmarks, and your document flow may behave differently.

In ChatGPT Work and Codex, the same subscription limit covers more tasks than with GPT-5.6 Terra, and substantially more than with Sol or Astra. For teams that keep hitting quotas, this is the fastest way to increase the volume of work without changing plans.