GPT-6.1 Terra Review: Agent Work at Half Sol's Price

GPT-6.1 Terra is OpenAI's new mid-tier model, the successor to GPT-5.6 Terra. OpenAI's launch post pitches it for computer use, visual reasoning, and writing, and says its standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol.

Most of OpenAI's claims are about cost per task on agent benchmarks, which we can't reproduce in a single prompt. So we priced the same agentic session on Terra and the four models OpenAI compares it against, then checked what each cost-per-task claim implies about token use.

The short version is that Terra costs half of Sol on every line of the rate card, and 37% of GPT-5.6 Terra on a cache-heavy session, because its cached input price fell 75%. Most of the "one-third the cost" figure against GPT-5.6 Terra comes from price alone. Against Claude Opus 5.5, OpenAI's two figures (three times the token efficiency, nine times lower cost) don't multiply out, which we cover below.

Price changes

Prices are from OpenAI's model pages for GPT-6.1 Terra, GPT-5.6 Terra, and GPT-6.1 Sol.

Against Sol, the 50% cut is exact on every line. Against GPT-5.6 Terra, it's 50% on input and cache writes but larger on output (58%) and cached input (75%). The cached input change comes from OpenAI's prompt caching guide, which bills cache reads at 0.1× input on most models and 0.05× on GPT-6.1 Sol and GPT-6.1 Terra.

Terra has a 1,050,000-token context window, with up to 922,000 input tokens and 128,000 output tokens. It accepts text and image input and returns text.

Example session

We used the same session shape as our Claude Opus 5.5 review and GPT-6 Sol and Luna review, so the numbers compare across all three posts.

Assumptions:

• 80 turns
• 60,000 cached input tokens per turn (files, tool definitions, screenshots, and history)
• 1,000 fresh input tokens per turn (new instructions)
• 2,000 cache-write tokens per turn (new context added to the cache), billed at 1.25× input
• 1,000 output tokens per turn on every model. We held this constant to isolate price.

Across the session, that comes to 4.8M cached tokens, 80,000 input tokens, 160,000 cache-write tokens, and 80,000 output tokens.

Claude Opus 5.5 prices are from Anthropic's pricing page, using the 5-minute cache write ($5 per million) and cache read ($0.20 per million). GPT-6 Astra prices are from its model page.

Against GPT-5.6 Terra, the largest single saving is cached input, $0.96 down to $0.24. On this session, cached input was 39% of GPT-5.6 Terra's bill and took the 75% cut. That's why the session saves 63% when the input price only fell 50%.

Against Sol, the result is 50% no matter how the session is shaped, since every rate is exactly half.

Against Opus 5.5 and Astra, the gap is wider on cache-heavy work. Terra's cached input is a quarter of Opus 5.5's and a twentieth of Astra's, while its input and output prices are a quarter and a tenth.

OpenAI's cost-per-task claims

OpenAI's launch post makes four comparisons on cost. We checked each against the rate cards to see how much comes from price and how much would have to come from Terra using fewer tokens.

Against GPT-5.6 Terra: "about one-third the cost." On Agents' Last Exam, a benchmark of professional computer-use tasks, Terra at medium effort scores 15 points higher than GPT-5.6 Terra at about a third of the cost. Price alone gets most of the way there. Our cache-heavy session costs 37% as much on Terra with identical tokens, which leaves about 10% fewer tokens to account for. With no caching, price alone gives 42% to 50% depending on the input and output mix, so Terra would need roughly 20% to 33% fewer tokens to reach a third. OpenAI doesn't publish the token counts, so we can't tell which case applies.

Against GPT-6 Astra: "on par" at one-tenth the price per token. OpenAI says Terra matches Astra on Agents' Last Exam and completes tasks over three times faster. It doesn't give a cost per task for this comparison. On our session with identical tokens, Terra costs 8% of Astra, or about 12.6 times less. The figure is lower than one-tenth because Terra's cached input is a twentieth of Astra's.

Against Claude Opus 5.5: "~9x lower cost per task." At max effort, OpenAI says Terra is over three times as token efficient as Opus 5.5 at a quarter of its price per token. Terra's rates are exactly a quarter of Opus 5.5's on input, cached input, cache writes, and output, so the input and output mix doesn't change the price ratio. Three times fewer tokens at a quarter of the price would be 12 times lower cost. A 9 times lower cost implies Terra uses about 2.25 times fewer tokens. OpenAI doesn't say which tokens its efficiency figure counts. If it counts only output tokens, the difference could come from input. We can't settle it from the published figures.

Against Opus 5.5 on science tasks. On Terminal-Bench-Science 0.1, OpenAI says Terra at max effort comes within five points of Opus 5.5 at less than a fifth of the estimated cost, and surpasses GPT-6.1 Sol at lower cost.

A few caveats apply to all four. Effort levels differ between claims, from medium on Agents' Last Exam against GPT-5.6 Terra to max against Opus 5.5. OpenAI publishes no absolute Agents' Last Exam scores, only the gaps. And these are OpenAI's own benchmarks, which we didn't rerun.

Capability claims

The other numbers in the launch post are about results rather than cost:

• Coding. Terra solves 50.3% of Terminal-Bench 4.0 tasks at max effort, against 21.5% for GPT-5.6 Terra. OpenAI says GPT-6.1 Sol and GPT-6 Astra "remain the better choice" for the most complex coding tasks.
• Desktop apps. On OSWorld 2.1 at medium effort, Terra surpasses GPT-5.6 Terra's best score at a fraction of the cost per task. No scores are given.
• Visual reasoning. On MMMU-Pro, Terra outperforms GPT-5.6 Terra at every reasoning setting and nearly matches Astra and Sol. No scores are given.
• Factual errors. On OpenAI's internal set of error-inducing prompts, Terra's error rate at low effort is 16.7%, against 53.6% for GPT-5.6 Terra. At max effort it's 4.3%. The system card describes a separate evaluation built from user-flagged ChatGPT conversations and says the results shouldn't be read as production hallucination rates.

Where Terra fits

OpenAI's lineup now has three price points for this kind of work. GPT-6 Luna costs $0.10 and $0.50 per million tokens, a tenth of Terra, and is aimed at summarization, extraction, and quick questions. Our session costs $0.116 on Luna, per our Sol and Luna review. Terra costs $0.92 and is aimed at computer use, visual tasks, and everyday agent work. Sol costs $1.84 and is OpenAI's pick, with Astra, for the hardest coding.

Because Terra is exactly half of Sol, Terra is cheaper as long as it uses fewer than twice Sol's tokens on a task. The case for Sol is success rate, since a failed Terra run that has to be retried can cost more than one Sol run.

Fast mode, batch, and long prompts

A few pricing options change these totals:

• Fast mode costs 2× standard rates, which puts Terra at $2 and $10 per million, the same as Sol at standard speed.
• Batch and Flex cost 50% less than standard.
• Prompts over 272K input tokens are billed at 2× the input and cache rates and 1.5× the output rate, for the entire request. Our session stays well under that at 63,000 tokens per turn. A long session that grows past 272K of context per turn will cost more per turn from that point on.

Checking it yourself

Every response reports its token usage, so you can replace our fixed 1,000 output tokens per turn with figures from your own tasks. Puter.js gives free access to GPT-6.1 Terra without an API key, and the same call works for gpt-5.6-terra and gpt-6.1-sol.

Script tag:

<html>
<body>
  <script src="https://js.puter.com/v2/"></script>
  <script>
    const prompt = `List the vendor, date, line items, and total from this invoice as JSON:

Northwind Supply, Invoice #4471, March 3, 2026
12 x USB-C cables at $8.50
4 x Monitor arms at $64.00
Shipping $15.00`;

    for (const model of ["gpt-6.1-terra", "gpt-5.6-terra", "gpt-6.1-sol"]) {
      puter.ai.chat(prompt, { model }).then(response => {
        puter.print(`${model}: ${response.message.content}<br><br>`);
        console.log(model, response.usage);
      });
    }
  </script>
</body>
</html>

npm (@heyputer/puter.js):

import { puter } from '@heyputer/puter.js';

const prompt = `List the vendor, date, line items, and total from this invoice as JSON:

Northwind Supply, Invoice #4471, March 3, 2026
12 x USB-C cables at $8.50
4 x Monitor arms at $64.00
Shipping $15.00`;

for (const model of ['gpt-6.1-terra', 'gpt-5.6-terra', 'gpt-6.1-sol']) {
  const response = await puter.ai.chat(prompt, { model });
  console.log(model, response.message.content);
  console.log(model, response.usage);
}

Run each task several times per model and average the output tokens, since length varies between runs. Multiply the averages by the rates in the table above to get your own cost per task. If you'd rather use the official SDK, Puter also exposes an OpenAI-compatible endpoint that returns the same usage fields.

What this adds up to

On price, OpenAI's 50% figure holds exactly against GPT-6.1 Sol and understates the cut against GPT-5.6 Terra. On our cache-heavy session, Terra costs 37% of GPT-5.6 Terra, because cached input fell 75% and output fell 58%.

That covers most of the "one-third the cost" claim against GPT-5.6 Terra before Terra uses any fewer tokens. Against Astra, identical tokens already put Terra at about a twelfth of the cost. Against Opus 5.5, the price ratio is a flat 4× on every token type, so OpenAI's 9× figure implies about 2.25 times fewer tokens, not the three times stated alongside it.

What the rate card can't show is whether Terra finishes the task. OpenAI's benchmarks say it does on most computer-use and visual work, and that Sol and Astra are still the better choice for the hardest coding. If your tasks fall in the first group, the price difference is large enough that Terra is worth testing first.

FAQ

What is GPT-6.1 Terra? GPT-6.1 Terra is OpenAI's mid-tier model in the GPT-6.1 family and the successor to GPT-5.6 Terra. OpenAI positions it for computer use, visual reasoning, and writing, with GPT-6.1 Sol and GPT-6 Astra above it and GPT-6 Luna below it.

How much does GPT-6.1 Terra cost? GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens. Cached input is $0.05 per million and 30-minute cache writes are $1.25 per million. Prompts over 272K input tokens cost more, and Fast mode costs twice the standard rate.

Is GPT-6.1 Terra cheaper than GPT-6.1 Sol? GPT-6.1 Terra costs exactly half of GPT-6.1 Sol on input, cached input, cache writes, and output. OpenAI says Terra completes Agents' Last Exam tasks over three times faster than Sol, and recommends Sol and GPT-6 Astra for the most complex coding tasks.

What is the GPT-6.1 Terra context window? GPT-6.1 Terra has a 1,050,000-token context window, with up to 922,000 input tokens and 128,000 output tokens. It accepts text and image input, and its knowledge cutoff is June 1, 2026.

Where can I use GPT-6.1 Terra? GPT-6.1 Terra is available in the OpenAI API as gpt-6.1-terra, and in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. Developers can also call it through Puter.js with the model ID gpt-6.1-terra.

Related

• OpenAI: Introducing GPT-6.1 Terra
• GPT-6 Sol and Luna review
• GPT-6 Astra review
• Claude Opus 5.5 review
• OpenAI API pricing: full breakdown of costs
• Free, unlimited OpenAI API with Puter.js
• Puter.js AI documentation