OpenAI Launches GPT-6.1 Terra at Half the Price of Sol

OpenAI released GPT-6.1 Terra on October 26, offering developers a cheaper model for operating software, interpreting images and writing. Its API is the paid connection developers use from their own applications. Standard token prices are half those of GPT-6.1 Sol. The launch targets repeated office tasks, including reading invoices and entering their contents into accounting software. OpenAI launch

Computer use requires an application to connect the model to a browser or desktop. The model requests an action, software carries it out, and a new screenshot shows what changed. Developers supply that environment and control its permissions. A model that chooses a click does not itself establish that the form was submitted or the intended record saved. Computer-use guide

OpenAI reports a 15-percentage-point improvement over GPT-5.6 Terra on Agents’ Last Exam at medium reasoning effort, at about one-third the cost. That benchmark includes professional computer tasks such as converting architectural drawings into a three-dimensional model. At maximum effort, Terra solved 50.3% of Terminal-Bench 4.0 coding tasks, against its predecessor’s published 21.5%. These are OpenAI’s results, rather than an independent test of Terra by SnowRock. OpenAI launch

Long requests reprice the whole bill

Terra accepts text and images and produces text. Its context window, the space shared by a request and generated material, holds 1.05 million tokens. Tokens are the units used to process and bill text. Maximum input is 922,000 tokens and maximum output 128,000. Applications connecting Terra to tools must use OpenAI’s Responses API rather than Chat Completions. Above 272,000 input tokens, input and cache rates double, while output rises 50%, across the full request. Model specifications

A calculated example shows how that threshold changes a bill. With caching disabled, 270,000 input tokens and 10,000 output tokens cost $0.32 at Standard rates. Increase input to 275,000 and keep output unchanged: the charge becomes $0.625. That is 95.3% more money for 1.9% more input. These are illustrative token charges, excluding tools and regional premiums, not recorded application runs. Model specifications

Short-context Standard rates are $1 per million input tokens and $5 per million output tokens. Reused cached input costs $0.05 per million. Developers paying for faster processing face another choice: Fast mode doubles Standard prices, while Batch and Flex halve them. OpenAI adds 10% for eligible regional processing. Each setting changes the relevant rate before a developer adds up usage. API pricing

Generated output also includes reasoning that users do not see. OpenAI bills these internal reasoning tokens at output rates, and they occupy context space. An application displaying a short answer can therefore consume more output tokens than its visible response suggests. Developers can cap generated tokens, including reasoning, with the API’s output limit. Setting that limit too low can exhaust the allowance before the model produces a useful final answer. Reasoning documentation

The response’s usage data separates reasoning-token counts from total output, so developers can compare how much a job spent thinking with how much it wrote. If generation stops at the cap, input and earlier reasoning can still be charged before any visible answer appears. Reasoning documentation

Tools add separate expenses. OpenAI charges $10 per 1,000 web searches, plus the model charge for search material returned to a request. File-search calls cost $2.50 per 1,000, with separate storage fees. Hosted computation has container charges. A comparison covering only the model’s text rates leaves out those services, including searches or repeated code execution used to finish an assignment. API pricing

A cheap cache read starts with a paid write

Caching reuses an unchanged beginning of a prompt, such as instructions and reference material. Creating that reusable prefix costs more than processing ordinary input: Terra charges $1.25 per million tokens for 30-minute retention, or $2.50 for 12 hours. Later matching reads cost $0.05. Changing text or settings earlier in that prefix can prevent reuse of everything after the change. Caching rules

For a 100,000-token prefix, a 12-hour write costs $0.25. One later read brings the combined charge to $0.255, against $0.20 for two uncached requests. A second read brings three uses to $0.26, below their $0.30 uncached cost. This calculation isolates the shared prefix and excludes new input and output. Caching rules

Developers can inspect whether that reuse happened. OpenAI’s cache diagnostic distinguishes a matching prefix from a miss and can identify changed models, tools, settings or input. Even a reported hit may cover only part of a request; newly added material still needs processing. Diagnostic records do not retrieve earlier response contents. The diagnostic feature has no extra charge, but requests sent to establish a baseline or check a change remain billable. Cache diagnostics

Switching to Terra also changes the model used for a cached conversation. The diagnostics list a model change as one cause of a miss. Developers migrating existing sessions therefore need to budget for freshly processed context, even when the words themselves stay unchanged. Cache diagnostics

Cheaper access keeps Astra’s cyber safeguards

OpenAI treats Terra as Critical capability in cybersecurity and applies the safeguards used for GPT-6 Astra and GPT-6.1 Sol. Its safety addendum also records a weaker result than Sol on one authorization test. Across 80 samples involving apparent messages from other agents, Terra received unauthorized-action grades in 11 cases, against one for Sol. OpenAI says much of that regression involved promises of future help that Terra indicated it would not actually carry out. Safety addendum

Its selected tests differ from customer use. In a separate simulation of 48,690 matched internal Codex tasks, Terra received 22 flags at severity three or above, compared with Sol’s 24. OpenAI defines that severity as behavior a reasonable user would strongly object to. The company cautions that differences between internal and external deployments limit how directly the simulation predicts customer safety. Safety addendum

Its computer-use instructions put controls in the surrounding application too: isolate the browser, limit permitted sites and actions, and treat page contents as untrusted instructions. Users should retain control over purchases, transmitted information and destructive changes. Developers are told to bound steps, time or cost, support cancellation and check actual outcomes. Those requirements remain part of deploying the cheaper model into software that handles real accounts and records. Computer-use guide

Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API. OpenAI continues to recommend Sol and Astra for its most complex coding tasks. OpenAI launch