OpenAI Launches GPT-6.1 Terra at $1 Input and $5 Output per Million Tokens

GPT-6.1 Terra launched on October 26

OpenAI published GPT-6.1 Terra on October 26, 2026. The official feed describes the launch as “fast and capable computer use,” while the live developer reference positions the model for matching product images to catalogs, interpreting technical drawings, acting on application screens, and drafting or revising text. The API identifier is gpt-6.1-terra.

The model reference is live and lists API rate limits for Build, Launch and Grow usage tiers; the Free tier is not supported. The canonical announcement page was blocked by a Cloudflare challenge during this review. AccessAllGPT therefore did not verify or repeat any launch-page benchmark, safety result, ChatGPT plan, geographic rollout or account-access claim. Builders should confirm the model appears in their own project before planning a cutover.

Standard pricing starts at $1 in and $5 out

For short-context Standard requests, OpenAI lists $1 per million input tokens, $0.05 per million cached-input tokens, $1.25 per million 30-minute cache-write tokens and $5 per million output tokens. A 12-hour cache write is $2.50 per million tokens. Batch and Flex are 50% below Standard, Fast is twice Standard, and regional processing adds 10% where available.

Those are token rates, not cost-per-completed-task results. Computer use can add tool-specific charges, screenshots and action history can enlarge prompts, and failed trajectories can create retries. A fair comparison must count model tokens, cache writes and reads, tool fees, retries, elapsed time, human intervention and the share of tasks that finish acceptably.

The 1.05M context has a price boundary at 272K input tokens

Terra accepts text and images and produces text. OpenAI documents a 1,050,000-token context window, up to 922,000 input tokens and 128,000 output tokens, plus a June 1, 2026 knowledge cutoff. Audio and video are not supported input modalities.

When a prompt exceeds 272,000 input tokens, OpenAI prices the full request at twice the input and cache rates and 1.5 times the output rate. Standard long-context rates are therefore $2 input, $0.10 cached input, $2.50 for a 30-minute cache write, $5 for a 12-hour cache write and $7.50 output per million tokens. The larger window is capacity, not evidence that every screenshot, document or instruction will be used reliably.

Computer use and other tools require Responses

OpenAI lists Chat Completions, Responses and Batch as supported endpoints, but says tool calling requires the Responses API. Supported Responses tools include computer use, web search, file search, image generation, Code Interpreter, hosted shell, apply patch, skills, MCP and tool search. Streaming, function calling, structured outputs and prompt caching are supported; fine-tuning and predicted outputs are not.

Reasoning effort supports none, low, medium, high, xhigh and max, with medium as the default. Pin that setting during evaluation: a test that leaves effort implicit is testing medium. Also pin the model alias, because the model page currently lists gpt-6.1-terra as both the default snapshot and the only snapshot rather than exposing a dated version.

The release does not yet establish computer-use reliability

The accessible first-party sources describe intended capabilities but do not provide a verified success rate for a reader’s applications. AccessAllGPT did not run Terra, and this article makes no claim that it surpasses Sol, Astra, Luna or a specialist computer-use model. “Precise computer use” remains vendor positioning until measured against representative screens, state changes and failure costs.

Evaluate complete trajectories, not isolated screenshots. Include ambiguous controls, changed layouts, loading states, permission prompts, inaccessible elements, stale observations, interrupted sessions and deceptive page content. Score correct completion, wrong action, unnecessary action, recovery, human intervention, latency and total cost. Preserve screenshots, model outputs, proposed actions, policy decisions and resulting state.

What AI teams should do next

Trial Terra where vision-heavy or interface-driven tasks make its price and positioning relevant. Constrain it to a test environment first, allow only the applications and actions named by policy, and require approval for purchases, messages, deletion, permission changes, credential handling and any irreversible state change. Keep deterministic validation between the model’s proposal and the executor.

Wait if account access, regional processing, tool pricing, long-context economics or recovery behavior is unresolved. Reject a production migration if Terra misses workload-specific quality or action-safety gates after retries are counted. The immediate decision is whether the model earns a bounded evaluation—not whether a launch label is enough to entrust it with a desktop.

Disclosures

AccessAllGPT did not receive OpenAI credentials, credits, early access, a briefing, benchmark data, review or compensation for this article. OpenAI did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI. Publication-wide relationships are listed on the disclosures page.