GPT-6.1 Terra: What's New, Benchmarks and How It Compares to Sol

GPT-6.1 Terra is OpenAI's fast, efficient middle-tier model, released October 26, 2026 as gpt-6.1-terra. Its API price is $1 per million input tokens and $5 per million output tokens — half of GPT-6.1 Sol, below GPT-5.6 Terra's $2 and $12 — and OpenAI reports large gains over GPT-5.6 Terra in computer use, coding and factual accuracy. For the most complex coding, OpenAI still points to Sol and Astra.

Short answer

• Model: gpt-6.1-terra, released October 26, 2026, per OpenAI's API changelog.
• Price: $1 input, $0.05 cached input and $5 output per 1M tokens, from OpenAI's Terra model page, checked October 27, 2026.
• Limits: 1,050,000-token context window, 128,000 max output tokens, text and images in, text out.
• Best for: visual reasoning, precise computer use and clear, natural writing, in OpenAI's words, plus high-volume work such as reading invoices and entering their details into accounting software.
• Where: the API, ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and GitHub Copilot.

What is GPT-6.1 Terra?

OpenAI's GPT-6 guide puts Terra in the third slot of a four-model lineup. Astra is the top tier, "for the most demanding reasoning, coding, and professional work". Sol is the balanced tier, near Astra at a lower cost. Terra is the fast and efficient tier, built for visual reasoning, precise computer use and clear, natural writing. Luna is the fastest and cheapest, for focused, high-volume tasks.

The launch post lists drafting proposals, interpreting technical drawings, filling out forms and editing videos as the kind of work Terra handles (announcement). Its standard rates are half of GPT-6.1 Sol's $2 and $10, and below GPT-5.6 Terra's $2 and $12.

Standard prices in US dollars per 1M tokens from OpenAI's model pages for Terra, Sol, GPT-5.6 Terra, Astra and Luna, checked October 27, 2026.

Long prompts cost more. Above 272K input tokens, Terra charges $2 input, $0.10 cached input and $7.50 output for the whole request. Batch and Flex cost 50% less than Standard, and Fast mode costs twice as much (model page).

What changed from GPT-5.6 Terra?

The price is the headline: input halves, output falls by 58%, and a cache read costs a quarter of what it did. The window and the output ceiling stay where they were, so existing prompts keep their shape. The knowledge cutoff moves forward by three and a half months, to June 1, 2026 (Terra, GPT-5.6 Terra).

Factual accuracy moves too. OpenAI says the error rate on its internal set of error-inducing prompts fell from 53.6% at low effort with GPT-5.6 Terra to 16.7% with Terra, a drop of nearly 37 percentage points. At maximum effort the rate is 4.3%, 18 points lower, nearly matching GPT-6 Astra and GPT-6.1 Sol (announcement). In ChatGPT Work and Codex, OpenAI says the same usage allowance now covers more tasks than it did with GPT-5.6 Terra.

GPT-6.1 Terra benchmarks

OpenAI publishes its results as prose in the launch post, and the charts on that page are images. These are OpenAI's numbers on OpenAI's test setups; we have not run Terra ourselves.

• Agents' Last Exam (effort: medium): 15 points above GPT-5.6 Terra at about one-third the cost; on par with GPT-6 Astra at one-tenth its price per token.
• Agents' Last Exam (effort: max): Over 3x more token-efficient than Claude Opus 5.5, one-fourth its price per token, about 9x lower cost per task.
• OSWorld 2.1 (effort: medium): Above GPT-5.6 Terra's best score at a fraction of the cost per task.
• MMMU-Pro (all effort levels): Beats GPT-5.6 Terra at every setting; nearly matches GPT-6 Astra and GPT-6.1 Sol.
• Terminal-Bench 4.0 (effort: max): 50.3% vs 21.5% for GPT-5.6 Terra.
• Terminal-Bench-Science 0.1 (effort: max): Above GPT-6.1 Sol at lower cost; within 5 points of Claude Opus 5.5 at under one-fifth the estimated cost.
• Error-inducing prompts (effort: low): 16.7% error rate, was 53.6%.

OpenAI also says Terra completes Agents' Last Exam tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol, and that for the most complex coding tasks Sol and Astra remain the better choice. That last line is the one to plan around: the price cut buys throughput, and the top of the lineup is still where the hardest work goes.

Independent results

Artificial Analysis measured Terra at 48 on its Intelligence Index at max effort, four points below GPT-6.1 Sol at 52. On its Coding Agent Index the gap widens, 48 against 54, with GPT-6 Luna at 31. Its output speed came in at 90.9 tokens per second for Terra against 69.4 for Sol, and it put cost per task at about $0.40, against $1.40 for GPT-5.6 Terra on its evaluations. On its hallucination test, Terra guessed 38% of the time when it did not know the answer, against 88% for GPT-5.6 Terra (Terra page, Sol page, launch analysis, checked October 27, 2026).

GitHub, which shipped Terra in Copilot the same day, says its own early testing had Terra solving noticeably more terminal and C# coding tasks than GPT-5.6 Terra. GitHub published no numbers with that claim (GitHub changelog).

GPT-6.1 Terra vs GPT-6.1 Sol

The two models share a window, an output ceiling and an input format. They differ on price, effort range, cutoff and measured intelligence.

Pick Terra for high-volume visual, computer-use and writing work where the half price matters. Pick Sol for the hardest coding, as OpenAI itself advises (announcement).

For a task-by-task comparison, read GPT-6.1 Terra vs GPT-6.1 Sol.

Where can you use GPT-6.1 Terra?

• OpenAI API: call gpt-6.1-terra on Chat Completions, the Responses API or Batch. Tool calling, including computer use, needs the Responses API (model page).
• ChatGPT Work and Codex: Plus, Pro, Enterprise and Edu users, from launch day (announcement).
• GitHub Copilot: generally available October 26, 2026 for Copilot Pro, Pro+, Max, Business and Enterprise, billed at provider list price (GitHub changelog).
• OmniaKey: OmniaKey does not offer GPT-6.1 Terra yet (checked October 27, 2026); GPT-6.1 Sol is available today with the same OpenAI API format.

Common questions

Is GPT-6.1 Terra better than GPT-6.1 Sol?

Not overall. OpenAI keeps Sol for the hardest coding, and Artificial Analysis scores Sol 52 against Terra's 48 on its Intelligence Index. Terra wins on price at half of Sol's per-token rate, and on Terminal-Bench-Science 0.1, where OpenAI reports Terra above Sol at lower cost.

Is a high-effort Terra better than a low-effort Sol?

A Hacker News reader asked this on launch day, and OpenAI publishes no head-to-head at mixed effort settings. Terra's per-token price is half of Sol's, but cost per task depends on how many tokens each setting burns. Run both on a handful of your own tasks and compare finished work per dollar.

Can GPT-6.1 Terra read images or make video?

It reads text and images and answers in text; OpenAI's model page lists audio and video as unsupported (Terra model page). The launch post's reference to editing videos describes Terra operating video software through computer use, with text in and text out.

Is GPT-6.1 Terra safe to use for security work?

OpenAI says Terra meets the Critical threshold in cybersecurity under its Preparedness Framework and keeps the protections built for GPT-6 Astra (announcement). The system card linked from that post carries the detail.

Update log

• October 27, 2026: first version, with OpenAI's October 26 announcement and model pages.