OpenAI releases GPT-6.1 Terra: API price cut in half to $1 per million tokens

OpenAI has released GPT-6.1 Terra, a mid-tier model for agentic workflows, computer use, and everyday coding. In the API it costs $1 per million input tokens and $6 per million output tokens. GPT-5.6 Terra was priced at $2 and $12, so the price has dropped exactly by half. By OpenAI's own estimates, the model is also noticeably stronger: on Terminal-Bench 4.0 it solves 50.3% of tasks versus 21.5% for its predecessor.

Cheaper per token and per task

Cutting the price of a mid-tier model usually means paying for the savings with quality. Here, OpenAI claims the opposite. On the Agents’ Last Exam benchmark, GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra at the medium reasoning level. The cost of completing a single task is roughly three times lower.

Simple arithmetic shows where the effect comes from. The rate fell by a factor of two, while the cost per task fell by a factor of three. That means the new model uses about 1.5 times fewer tokens for the same work: fewer redundant steps and less empty reasoning. For agents this matters more than the per-token price, because they run long chains of calls, and the bill grows with the number of steps.

What OpenAI's own tests show

All figures below come from OpenAI itself; the release includes no independent measurements.

• Input token price: $1 per million instead of $2 for GPT-5.6 Terra.
• Output token price: $6 per million instead of $12.
• Agents’ Last Exam: 15 percentage points better than the predecessor at the medium reasoning level, with the cost per task roughly three times lower.
• Terminal-Bench 4.0: 50.3% of tasks solved versus 21.5% for GPT-5.6 Terra at the maximum reasoning level. The gap is 28.8 percentage points, a 2.3-fold improvement.
• Factual errors: in an internal test on difficult, provocative prompts at the maximum reasoning level, their share fell to 4.3%.
• Intended use: agentic workflows, computer use, everyday coding.

Terminal-Bench tests whether a model can work in a command line: install dependencies, fix builds, and make sense of an environment. A jump from 21.5% to 50.3% on such a test means the agent gets stuck on routine work less often, even though routine work makes up most of real-world development. The drop in the error rate to 4.3% also matters for agents. A person will catch a made-up fact in a chat, but an agent will execute a made-up file path or command flag.

Where the model is available

GPT-6.1 Terra is already available in ChatGPT Work, Codex, and through the API under the identifier gpt-6.1-terra. In ChatGPT it is rolling out to the Plus, Pro, Enterprise, and Edu plans. Free users are not on the list.

How to recalculate your agent budget

Teams that run agents in production should re-run their own tasks on gpt-6.1-terra and compare the results with GPT-5.6 Terra. The thing to watch is not the per-token price but the total cost of a completed task and the share of successful runs. For a typical request of one million input tokens and one hundred thousand output tokens, the bill drops from $3.2 to $1.6 from the rate cut alone. If the reduction in tokens per task is also confirmed, the savings will be noticeably larger.

Reasoning modes deserve a separate check. OpenAI measured the Agents’ Last Exam gain at the medium level, while the Terminal-Bench record and the 4.3% error rate were obtained at the maximum level. The level you choose will set the balance between speed, cost, and quality. For developers in Codex, the model becomes a cheap default for routine coding, and heavier models can be saved for complex architectural tasks.