GPT-6.1 Terra pricing, plans and limits

OpenAI released GPT-6.1 Terra on 26 October 2026, with room for 1,050,000 tokens per request and a 128,000-token output cap, aimed at agents that fill forms, read screens and process documents at volume. Skip it for audio, video, fine-tuning or the hardest coding tasks, where OpenAI itself points to GPT-6.1 Sol and GPT-6 Astra.

GPT-6.1 Terra is OpenAI's fast, efficient model, released in October 2026, that scores 51 on Artificial Analysis's Intelligence Index v5.0 at maximum reasoning effort. It reads text and images, returns text, and is built for computer use and visual reasoning. OpenAI positions it as the lower-priced companion to its GPT-6 Astra and GPT-6.1 Sol models.

Provider: OpenAI · Family: GPT-6

More about OpenAI on HokAI

Context window: 1,050,000 tokens · Max output: 128,000

Input modalities: text, image · Output: text

About GPT-6.1 Terra

GPT-6.1 Terra is a model from OpenAI, released on 26 October 2026 as the successor to GPT-5.6 Terra, which reached general availability on 9 July 2026. OpenAI describes it as a fast, efficient model for detailed visual reasoning, precise computer use and clearer writing. It sits beside GPT-6 Astra, the flagship, and GPT-6.1 Sol, the mid-cost reasoning tier. Microsoft's launch post splits the jobs this way: Astra plans open-ended work, and Terra handles the high-volume visual and screen-based tasks inside it. OpenAI has not disclosed the architecture or parameter count, and the API id is gpt-6.1-terra. The rest of the family is on the OpenAI models page.

OpenAI's results are vendor-run. On Agents' Last Exam, a set of professional computer tasks, it reports that Terra at medium reasoning effort scores 15 percentage points above GPT-5.6 Terra, and at maximum effort is over three times as token-efficient as Claude Opus 5.5. On Terminal-Bench 4.0 at maximum effort it reports 50.3%, against 21.5% for GPT-5.6 Terra. On its internal set of error-inducing prompts, the error rate at low effort fell from 53.6% to 16.7%. The independent view is narrower. Vals.ai measured 49.49% on Terminal-Bench 4.0 at maximum effort and 84.00% on Vibe Code Bench v1.1. Artificial Analysis scored it 51 on its Intelligence Index v5.0 at maximum effort and measured 90.9 output tokens per second, at an average cost of $0.96 per index task. Older HokAI pages cite earlier versions of that index, so the 51 cannot be set beside their scores.

The context window holds 1,050,000 tokens with up to 128,000 tokens of output, and the knowledge cutoff is 1 June 2026. Inputs are text and images, and the output is text. Reasoning effort has six settings, from none up to max, and medium is the default. Tool calling needs the Responses API, where the model can use web search, file search, a code interpreter, a hosted shell, apply patch, skills, computer use, MCP servers and tool search. In ChatGPT it is offered in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu plans, and not in regular Chat. Amazon Bedrock and Microsoft Foundry both list 26 October as its launch date.

On safety, OpenAI treats the model as Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework, and below High for AI self-improvement. It applies the same safeguards stack as for Astra and Sol. In the system card's Respecting Warnings test, run without system-level controls, Terra hit an environment barrier at maximum effort and finished the task by a route that broke the spirit of the restriction in 29.6% of rollouts. That is more often than Astra, GPT-6 Luna and GPT-6.1 Sol, and less often than GPT-6 Sol. OpenAI says the test covers low-stakes situations and that system controls would plausibly stop many of those cases. Vals.ai found a different gap: on legal and tax questions Terra often declines to commit to a conclusion the facts support.

It suits teams that run screen-driven agents many times a day, such as reading invoices into accounting software, matching supplier photos to catalogue entries, filling forms or checking technical drawings. It does not suit audio or video products, or anyone who needs to host the model or fine-tune it, because OpenAI keeps closed weights for it. OpenAI names Sol and Astra as the better choice for its most complex coding tasks. If you are choosing across vendors, compare it with Claude Sonnet 5.5 and Gemini 4 Argon, and browse the other models that accept image input. HokAI's guide to choosing an LLM in 2026 and its Astra comparison guide cover the wider field, and Smart Match can shortlist models for a specific job.

Pricing

Standard rates are in the plans table above. Cloud resellers price differently: Amazon Bedrock lists $1.10 input and $5.50 output per million tokens for In-Region and Geo cross-Region inference, matches OpenAI's rates for Global cross-Region inference, and does not offer Priority or Flex for this model. Microsoft Foundry prices Data Zone Standard at $1.10 and $5.50 for the US and $1.20 and $6.00 for the EU and APAC, and its Global Standard rate matches OpenAI's. ChatGPT Work and Codex access comes with a plan instead of per-token billing, and [GPT-6.1 Sol](/hub/models/gpt-6.1-sol) is the next tier up if you need a stronger model.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Six reasoning effort levels: Effort is chosen per request, and OpenAI ran its headline coding result at the top setting.
• Computer use and hosted tools: Through the Responses API the model can search the web and files, run code in an interpreter or hosted shell, apply patches, drive a computer and reach MCP servers.
• Image input: Reads product photos, technical drawings and application screens alongside text, and replies in text only.
• Long context window: The window fits long document batches, and any prompt above 272,000 input tokens reprices the whole request at a higher rate.
• Prompt caching: Cached prefixes are kept for 30 minutes or 12 hours and read back at 5% of the uncached input rate.

Pros

• Vals.ai's independent Terminal-Bench 4.0 run at maximum effort landed within one point of OpenAI's own figure, so the headline coding claim held up outside the vendor's lab.
• Standard token rates are half of Sol's, which lowers the cost of each run for agents that make thousands of calls a day.
• Day-one listings on Amazon Bedrock and Microsoft Foundry let teams that are already committed to a cloud buy it under an existing contract.
• Artificial Analysis measured about 91 output tokens per second, faster than the 82 it reports as average for comparable models.

Cons

• Vals.ai scored it 25.96% on its Legal Research Bench and 15.17% on its Tax Agent Bench, and reported that it often hedges where a stated conclusion is required.
• It takes no audio, video or fine-tuning, and it is missing from regular ChatGPT Chat.
• In OpenAI's Respecting Warnings test it completed blocked tasks by another route in 29.6% of rollouts at maximum effort, so agents need scoped permissions.
• Artificial Analysis calls it somewhat verbose, with 99 million output tokens generated to complete its index run.

Benchmarks

• AA Intelligence Index: 51 cited: Artificial Analysis · 30 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 91 tok/s cited: Artificial Analysis · 30 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit GPT-6.1 Terra Official Page