GPT-6.1 Terra and the price of a finished task

OpenAI’s October 26 release cuts listed token rates. Its API conditions and evaluation caveats explain why a cheaper token is only part of the calculation.

Reporting · 2026-10-30 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · OpenAI’s October 26 API changelog records the release of GPT-6.1 Terra under the identifier gpt-6.1-terra. The entry places it in the Responses and Chat Completions APIs and describes visual reasoning, computer use and writing. That dated record puts the model inside this edition’s October 23–29 reporting week. The practical question is what a lower listed price buys, and which parts of an automated job that price leaves unresolved. [1]

Primary-source statement · The company’s release notes separate three access routes. They describe Terra rolling out in the API, ChatGPT Work and Codex, and distinguish Work and Codex from ordinary Chat. They also say API usage is billed separately from ChatGPT subscriptions. Someone choosing a model within an existing subscription is therefore reading a different billing offer from a developer sending usage to the API. A rollout statement does not demonstrate access in every account. [2]

Primary-source statement · Those notes also identify an infrastructure limit: UAE inference residency is unsupported. OpenAI explains that this concerns where inference runs, rather than where the user is located. It is a narrower statement than saying users in a country cannot use the model. The document does not establish the data-location arrangement of any particular customer’s application. [2]

Primary-source statement · For standard API usage, OpenAI lists $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. Its launch account calls these rates 50 percent lower than GPT-5.6 Terra and GPT-6.1 Sol. A token rate alone does not measure the price of completing a drawing, form or invoice. [3]

Primary-source statement · An illustrative request with 100,000 ordinary input tokens and 20,000 output tokens would cost ten cents for input and another ten cents for output at those rates: twenty cents combined. This is arithmetic, not a recorded bill or completed-task test. It assumes those volumes and standard treatment, excluding separate tool charges. More requests or output would change the calculation. [3]

Primary-source statement · The API model page lists separate cache-write charges for 30-minute and 12-hour retention. For requests above 272,000 input tokens, it says input and cache rates double and output rates rise by half for the entire request. Quoting a standard price therefore requires its size and billing category. [4]

Primary-source statement · That page lists a 1,050,000-token context window and 128,000 maximum output tokens: admitting a large context does not promise an equally long answer. It lists text input and output, image input, and no audio or video support. These are the issuer’s interface specifications, not a completed application test. [4]

Primary-source statement · The documentation lists reasoning settings from none to max, with medium as the default, and requires the Responses API for tool calling. OpenAI’s launch comparisons name particular effort settings. A result reported at maximum effort is consequently not the measured result of a default request. The model identifier alone does not specify the full configuration. [4]

Primary-source statement · On Agents’ Last Exam, OpenAI reports a 15-percentage-point gain over GPT-5.6 Terra at medium effort and about one-third the cost. Its coding comparison uses maximum effort: 50.3 percent on Terminal-Bench 4.0, against a published 21.5 percent for GPT-5.6 Terra. These remain company evaluations with different settings. [3]

Primary-source statement · OpenAI’s system-card addendum qualifies how such results travel into products. It says evaluations ran in its research environment or through the API, and that system prompts, available tools and effort can produce differences from production ChatGPT. It separately warns that comparison values for earlier models may reflect later versions and differ from their original launch values. A historical model name does not guarantee an unchanged comparison specimen. [5]

Primary-source statement · The same addendum says OpenAI treats Terra as Critical for cybersecurity, High for biological and chemical capability, and below High for AI self-improvement under its Preparedness Framework. It says the model receives the safeguards stack used for GPT-6.1 Sol and GPT-6 Astra, with compatible probes. These are the issuer’s assessments and safeguards statements; they are not an external certification or a guarantee about an individual workflow. [5]

Primary-source statement · The dated release record answers whether there is a new offering to examine. It does not answer whether a particular organization’s form filling or document review will finish correctly, how many attempts it will take, or what the resulting bill will be. Those remain questions for measured work in the intended application. The release makes a named service and its stated terms available for evaluation; the outcome of that evaluation is still separate. [1]

Reporting note. OpenAI originals checked October 27. All five documents share one issuer origin. Performance and safety statements are attributed company assessments; no paid call, independent benchmark, installation or individual account-access test was performed.

Corrections

No corrections recorded.

Report an error