What Is GPT-6.1 Terra? Pricing, Performance, and How to Choose Between It, Sol, Astra and Claude

GPT-6.1 Terra is the mid-tier model OpenAI released on October 26, 2026. Its API price is $1 input and $5 output (per million tokens), half that of GPT-6.1 Sol. In OpenAI's own testing, it matched the higher-tier GPT-6 Astra on a benchmark that simulates professional work. On the other hand, GPT-6.1 Sol and GPT-6 Astra scored higher on a benchmark that measures computer use alone and on a hard coding benchmark, and OpenAI itself still recommends Sol and Astra for hard coding. For writing, OpenAI says it has become clearer and more natural, but it published no benchmark numbers. Based on the original text of OpenAI's announcement and model pages (checked October 28, 2026), this article covers pricing, performance and where you can use it, plus how to choose between it, Sol, Astra and Claude.

1. Where GPT-6.1 Terra fits

OpenAI's main API models come in four tiers, from top to bottom: Astra, Sol, Terra and Luna. Terra is third from the top and succeeds GPT-5.6 Terra. On its model page, OpenAI describes it as "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing."

Source: OpenAI's Models page and each model's page (checked October 28, 2026). All four have a context window of about 1.05 million tokens and a maximum output of 128,000 tokens.

With a knowledge cutoff of June 1, 2026, it is the newest of the four. Like Luna, it lets you choose reasoning effort starting from "none" (no reasoning). It accepts text and image input, but not audio or video. It does not support fine-tuning.

Note that the guidance at the top of the API model list still reads "If you're not sure where to start, use GPT-6 Astra, our flagship model for complex reasoning and coding. Choose GPT-6.1 Sol to balance intelligence and cost, or GPT-6 Luna for cost-sensitive, high-volume workloads." Terra is not part of that sentence (as of October 28). Terra appears as the fourth entry in the list of main models on the same page.

2. Pricing

US dollars per million tokens. Source: OpenAI's GPT-6.1 Terra and GPT-6.1 Sol pages (checked October 28, 2026). Processing with a specified region costs 10% more.

The unit prices are exactly half those of GPT-6.1 Sol and the previous GPT-5.6 Terra. Compared with Claude, that is half of Sonnet 5.5 ($2 input, $10 output) and a quarter of Opus 5.5 ($4, $20). However, the number of tokens a task consumes varies by model and reasoning effort, so the "cost per task" figures in the next section are a more reliable way to compare actual costs.

3. Performance: where it is strong and where it is weak

OpenAI's announcement includes six charts showing each model's score and cost per task at each reasoning effort. Here we list each model's value at its highest setting (max). All of these were measured by OpenAI, not by a third party.

Values at each model's max reasoning effort. The small figures are cost per task (USD). Source: charts in OpenAI's "Introducing GPT-6.1 Terra" (checked October 28, 2026). The Claude column includes only the models OpenAI put in its charts: Opus 5.5 appears in only two benchmarks, and Sonnet 5.5 only in two others. Only a single max data point is shown for Opus 5.5; for Opus 5.5 on Terminal-Bench-Science, the legend says "with fallback," and its cost is OpenAI's estimate (from the body of the announcement). OSWorld 2.1 values are for the offline set. Factual errors is an OpenAI internal evaluation using prompts designed to elicit errors.

Three takeaways stand out.

(1) On work tasks, it matches the higher-tier models. Agents' Last Exam measures professional tasks such as building a 3D model from architectural drawings or watching a film shot by shot to produce a log. At max, Terra scored 60.0%, above 6.1 Sol (58.6%) and 6 Astra (59.3%). Even at medium reasoning effort it scored 57.6%, the same as Astra at medium, at a cost of $0.29 per task (Astra: $4.82). Opus 5.5 is higher at 63.2%, but costs about 9 times as much.

(2) On coding and computer use, the higher-tier models win. Terminal-Bench 4.0 is a coding benchmark in which the model fixes problems using terminal commands. At max, Terra scored 50.3%, more than double the previous Terra (21.5%), but it falls short of 6.1 Sol (55.9%) and 6 Astra (58.2%). OpenAI also says that GPT-6.1 Sol and GPT-6 Astra remain better suited for the most complex coding tasks. On OSWorld 2.1, which measures computer use alone, 6.1 Sol scored 78.0% at max ($1.15 per task), beating Terra's 74.2% ($1.19) at almost the same cost, and Sol was also higher at medium and low. Claude Sonnet 5.5, which OpenAI included in its charts, scored 70.6% on Terminal-Bench 4.0 and 83.9% on OSWorld, the highest values in the table on both (at about 5 to 6 times Terra's cost).

The exception is the scientific code benchmark (Terminal-Bench-Science), where Terra scored 58.3% at max, beating 6.1 Sol (57.0%) at a lower cost. It still falls short of 6 Astra (68.1%) and Opus 5.5 (63.3%).

(3) Errors dropped sharply. On questions designed to elicit errors, the error rate at low reasoning fell from 53.6% for the previous Terra to 16.7%. At max it is 4.3%, almost the same as 6.1 Sol (4.4%) and 6 Astra (4.2%). On image understanding (MMMU-Pro), it scored 85.5%, within about 1 point of Sol and Astra.

4. Terra vs. Sol, Astra and Claude: which to use

Taking OpenAI's measurements at face value, the deciding factors are "what kind of work it is" and "how much you can spend per run."

• Running routine work tasks at volume, such as processing documents, emails and spreadsheets: Terra It ranks alongside the top models on the work-task benchmark (Agents' Last Exam), and its cost per task at max is about half of Sol's and about one-seventh of Astra's. OpenAI's announcement gives the example of reading incoming invoices and entering them into accounting software.
• Work centered on computer use: 6.1 Sol On OSWorld 2.1, Sol scored higher at about the same cost. Sol's model page also says to use Sol for complex coding, computer use and professional work.
• Hard coding and large changes: 6.1 Sol or 6 Astra Both outscore Terra on the terminal coding benchmark. OpenAI itself recommends this split.
• Scientific computing and the hardest problems: 6 Astra On Terminal-Bench-Science, Astra had the highest score at 68.1%. It costs about 6 times as much as Terra.
• If you use Claude In OpenAI's charts, Opus 5.5 scored about 3 points higher than Terra on work tasks, and Sonnet 5.5 scored highest on coding and computer use. Both cost more than Terra. Keep in mind, though, that this comparison uses benchmarks and settings OpenAI chose. For comparisons from the Claude side, see "Claude Sonnet 5.5: performance and pricing" and "Claude Opus 5.5: performance and pricing."

If you can't decide between Sol and Terra, the most reliable approach is to run the same task through both a few times and compare not the scores but "how much fixing the output needs" and the bill. Benchmark numbers are values measured on that benchmark with those settings. For differences from the previous generation, see also "GPT-6.1 Sol vs. GPT-6 Sol" and "our GPT-6 Astra guide."

5. Using it in ChatGPT and Codex

In ChatGPT, you can select GPT-6.1 Terra from the model list in ChatGPT Work and Codex. The announcement lists Plus, Pro, Enterprise and Edu as eligible, and the ChatGPT release notes (October 26) also include Business. It is rolling out gradually to accounts, and it may not be available depending on your plan, region or workspace settings.

• Usage limits themselves have not changed. The release notes say "overall plan usage limits are unchanged." On top of that, the announcement says you can get more work done within the same limits than with GPT-5.6 Terra, and much more than with 6.1 Sol or 6 Astra.
• Your default model may change. According to the release notes, for some Plus and Pro users, the default "Power" setting in ChatGPT Work may appear to have changed from "GPT-6.1 Sol Light" to "GPT-6.1 Terra Medium." If responses feel different than usual, check which model is displayed.
• You can also select it in Codex. The Codex CLI added GPT-6.1 Terra to its model list in 0.165.0 (October 28).

Sources: OpenAI's ChatGPT release notes (October 26, 2026 entry) and Codex 0.165.0. UAE inference residency is not supported (release notes).

For how usage limits differ across ChatGPT plans, see "ChatGPT Pro 100, 200 and 500 compared."

6. Safety

OpenAI announced that GPT-6.1 Terra reached the "Critical" threshold in the cybersecurity category of its safety framework (the Preparedness Framework). For that reason, it uses the same safeguards built for GPT-6 Astra. OpenAI also says that, compared with GPT-5.6 Terra, it is more likely to stay within the scope it was asked to work in and to honestly report constraints that block its work, and that it is more resistant to malicious instructions (prompt injection). Details are in OpenAI's System Card.

Summary

FAQ

Q. How much does the GPT-6.1 Terra API cost?

A. Per million tokens: $1 input, $0.05 cached input and $5 output. Batch and Flex are half price, and Fast is 2x. For input over 272,000 tokens, input and cache for the whole request cost 2x and output costs 1.5x.

Q. Should I use GPT-6.1 Sol or Terra?

A. Use Terra for running routine work tasks such as document processing at volume, and Sol for hard coding or work centered on computer use. In OpenAI's measurements, Terra beat Sol on the work-task benchmark, while Sol scored higher on the terminal coding and computer use benchmarks. Terra's unit price is half.

Q. Does it outperform Claude Opus 5.5?

A. On the two benchmarks OpenAI included, Opus 5.5 scores higher, while Terra costs much less (the Opus 5.5 cost on the scientific code benchmark is OpenAI's estimate). On the work-task benchmark, Opus 5.5 scored 63.2% ($8.82 per task) and Terra 60.0% ($1.01).

Q. Can I use it on the free ChatGPT plan?

A. The announcement and release notes cover Plus, Pro, Business, Enterprise and Edu; Free is not included. In the API usage tiers, Free is also listed as "not supported."

Q. Should I switch from GPT-5.6 Terra?

A. The unit price is halved, and in OpenAI's measurements it beats GPT-5.6 Terra on all five benchmarks where GPT-5.6 Terra appears, so there is little reason not to switch. In the API, change the model name to gpt-6.1-terra, and check the default reasoning effort (medium) and the fact that tool calling requires the Responses API. For the previous generation, see "our GPT-5.6 guide."

Update log

• October 28, 2026 First published. Written after checking OpenAI's announcement (October 26) and its chart values, the API model page and the ChatGPT release notes.

Sources

• OpenAI: Introducing GPT-6.1 Terra, GPT-6.1 Terra (API), Models, GPT-6.1 Sol, GPT-6 Astra, ChatGPT release notes, Codex 0.165.0
• Anthropic (comparison prices): Pricing