OpenAI introduced GPT-6.1 Terra on October 26, 2026, a model it positions between GPT-5.6 Terra and its flagship GPT-6 Astra, and set its standard API price at $1 per million input tokens and $5 per million output. It was live the same day for Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and for developers through the API as gpt-6.1-terra.
What was actually released
Terra is a vision-first model. OpenAI describes it as combining detailed visual reasoning and precise computer use with clearer writing, and names the work: drafting proposals, reading technical drawings, filling out forms and editing video. The page calls it a step up from GPT-5.6 Terra on factuality and coding, and keeps Sol and Astra above it for the most complex coding tasks.
What ships is a model with a rate card attached: $1 per million input tokens, $0.05 per million cached input and $5 per million output. OpenAI says the standard rates are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol, and that inside ChatGPT Work and Codex the same usage allowance now covers more tasks than it did with GPT-5.6 Terra.
Availability is immediate rather than staged: no waitlist, no preview period, no regional list. Every plan that already pays for ChatGPT gets the model, and the API identifier is gpt-6.1-terra.
The cached input rate is the line a production team should read twice. At $0.05 per million it is a twentieth of the standard rate, applied to the instructions and transcript prefix an agent resends on every call.
What the claim rests on
The page carries six figures, all of them produced by OpenAI.
| Claim on the page | Figure published | Baseline it is measured against |
|---|---|---|
| Computer use, Agents’ Last Exam, medium effort | 15 percentage points higher | GPT-5.6 Terra, at about one-third the cost |
| Computer use, Agents’ Last Exam, max effort | ~9x lower cost per task | Claude Opus 5.5 |
| Visual reasoning, MMMU-Pro | ahead at every effort setting | GPT-5.6 Terra, close to GPT-6 Astra |
| Desktop tasks, OSWorld 2.1, medium effort | above the previous best score | GPT-5.6 Terra, fraction of cost per task |
| Coding, Terminal-Bench 4.0, max effort | 50.3% of tasks solved | GPT-5.6 Terra at 21.5% |
| Error rate, low effort, internal benchmark | 16.7% | GPT-5.6 Terra at 53.6% |
Figures as published by OpenAI on October 26, 2026, reproduced as the company’s own runs. The Terminal-Bench 4.0 baseline is the previously published score for GPT-5.6 Terra, not a run made for this comparison.
The comparison with GPT-5.6 Terra is the one a reader can check against a second OpenAI page, and it holds. On July 30, 2026, OpenAI priced GPT-5.6 Terra at $2 per million input tokens and $12 per million output. Terra lands at $1 and $5, which is input down 50% as announced, output down 58%. The headline understates the output cut.
Against Claude Opus 5.5, the page says Terra costs one-fourth as much per token, and that figure matches the rate we have on record for Opus 5.5 at $4 input and $20 output. The cost per task claim is the weaker half of the same sentence. Three times the token efficiency at a quarter of the price works out to one twelfth of the cost, or twelve times cheaper, not the nine times announced. No token count per task is published, so the gap between the page’s own two numbers cannot be resolved from what was released.
At max effort, Terra is over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task.
What the page does not publish
No context window, no latency or throughput figures, no rate limits, no retirement date for GPT-5.6 Terra and no independent evaluation. The error benchmark is described only as internal, with no name and no prompt set behind it. Every row of the table above is an OpenAI run against an OpenAI baseline.
On this page
What it changes for teams paying for computer-use work
Terra arrives cheaper than the two frontier models beside it, on both sides of the meter.
| Model | Input per million | Output per million | Status of the rate |
|---|---|---|---|
| GPT-6.1 Terra | $1 | $5 | launched October 26, 2026 |
| Gemini 4 Argon | $2 | $10 | introductory, no end date |
| Claude Opus 5.5 | $4 | $20 | published list rate |
| Claude Haiku 5.5 | $0.10 | $0.50 | small model, rises above 100,000 tokens |
Rates as published by each vendor, read on the dates cited: OpenAI on October 26 and July 30, 2026, Google on October 13, 2026, Anthropic on October 8, 2026. The Argon rate is called introductory by Google and carries no end date; Haiku 5.5 input rises to $0.50 per million above a 100,000-token prompt.
The segmentation matters more than the discount. OpenAI states plainly that Sol and Astra remain the better choice for the hardest coding tasks while Terra takes the high-volume reading and clicking. That is a vendor telling buyers to route by task rather than by model generation, and it changes an agent budget: the expensive model keeps the planning step, and a cheap one absorbs the screenshot reading and extraction that dominate the token count.
Two earlier releases this month set the same test. Anthropic priced Haiku 5.5 at $0.10 per million tokens with an effort dial, and Google shipped Gemini 4 Argon at $2 per million tokens as an introductory rate with no end date. Terra joins them on that axis: the price of a task is set by the effort setting you pick, not by the badge on the model.
The Neural Minute take
The interesting number here is not a benchmark, it is the shape of the line. With Terra, OpenAI publishes a price one quarter of Opus 5.5 and one half of its own Sol, plus a paragraph telling you where not to use it. That is a company segmenting its catalog by effort level, and it moves the buyer’s decision from model to setting.
Our reservation is about the number a procurement team will quote. The 16.7% error rate is measured on prompts written to induce errors, so it describes resistance to traps, not accuracy on your invoices and shipping labels. A model that halved its mistake rate on a trap benchmark is still the wrong instrument for a workflow where one wrong digit costs a resubmission.
The practical move is to price one real task. Take a document type you process daily, run 50 of them at low, medium and max effort, and count finished items rather than tokens. Terra at $1 input makes the experiment cheap; the effort setting that clears your quality bar is the only rate that matters.
Sources
- Introducing GPT-6.1 Terra, OpenAI, published October 26, 2026, read on October 27, 2026: https://openai.com/index/introducing-gpt-6-1-terra/
- The date of the fact, October 26, 2026, is shown on the OpenAI news index and in OpenAI’s own feed for that page, not inferred from the feed that surfaced the item
- Advancing the price-performance frontier with GPT-5.6, OpenAI, published July 30, 2026, for the $2 and $12 per million rates of GPT-5.6 Terra: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6
- Figures quoted here, all from the announcement page: the three rates, the Agents’ Last Exam gap, the Terminal-Bench 4.0 pair and the two error rates in the table above
- Rates for the other models come from each vendor’s own page, read on the dates given in the caption