GPT-6.1 Terra vs GPT-6.1 Sol: Benchmarks, Speed and Which to Use

Use GPT-6.1 Terra for high-volume visual, computer-use and writing work: it costs half as much per token, $1 input and $5 output per million against $2 and $10 for Sol, and OpenAI and Artificial Analysis both report it faster. Use GPT-6.1 Sol for the hardest coding and reasoning: OpenAI says Sol and Astra remain the better choice for the most complex coding, and Artificial Analysis scores Sol higher, 52 against 48.

Short answer

The science row follows OpenAI's report on Terminal-Bench-Science 0.1: at max reasoning, Terra scores above GPT-6.1 Sol at lower cost.

Specs and prices side by side

Both model pages are OpenAI's, checked October 27, 2026: GPT-6.1 Terra and GPT-6.1 Sol. Prices are OpenAI's list prices in US dollars per million tokens.

Batch and Flex are 50% off for both, Fast mode is 2x, and Sol adds an Ultrafast mode at 6x, per OpenAI's pricing page (checked October 27, 2026).

Both accept text and images, return text and cap output at 128,000 tokens. Terra's model page puts its input ceiling at 922,000 tokens and lists computer use, code interpreter, hosted shell, apply_patch, MCP and tool search, plus streaming, structured outputs and web search; tool calling runs on the Responses API.

What a request costs

The sums below are arithmetic on OpenAI's list prices, not a bill from anyone's account.

100,000 input tokens plus 10,000 output tokens comes to $0.10 + $0.05 = $0.15 on Terra and $0.20 + $0.10 = $0.30 on Sol. At 400,000 input and 20,000 output the request is above 272K, so the long-context rates apply to the whole request: Terra pays $0.80 + $0.15 = $0.95 and Sol pays $1.60 + $0.30 = $1.90.

Cache reads run $0.05 per million on Terra against $0.10 on Sol. Terra's cache write costs $1.25 per million for 30 minutes and $2.50 for 12 hours; Sol's cache write is $2.50.

Price per token is half, but cost per task depends on how many tokens a run uses, and reasoning effort changes that. Artificial Analysis measured about $0.40 per task for Terra against $1.40 for GPT-5.6 Terra on its evaluations.

Benchmarks: what OpenAI reports

OpenAI's announcement gives its benchmark results in prose, and only a few of them compare Terra with Sol directly. Agents' Last Exam covers professional tasks such as turning architectural drawings into a 3D model and reviewing an animated film shot by shot to build a time-coded log for a restoration team. MMMU-Pro tests visual reasoning, OSWorld 2.1 tests desktop apps, Terminal-Bench 4.0 tests command-line coding, and Terminal-Bench-Science 0.1 tests science tasks carried out with code.

OpenAI then draws the line that matters for buyers: "For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." The company publishes its charts as images, and its text gives no Sol score for Terminal-Bench 4.0.

Benchmarks and speed: independent results

Artificial Analysis, an independent benchmarking site, scored both models at max effort on its Intelligence Index and Coding Agent Index and measured their output speed (Terra page, Sol page, launch analysis, checked October 27, 2026).

GitHub's early testing, reported in its Copilot changelog, found that Terra solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra. GitHub publishes no numbers with that claim. We have not run either model ourselves.

Which one should you use?

Documents, forms and computer use

OpenAI built Terra for visual reasoning and precise computer use: drafting proposals, reading technical drawings, filling in forms, entering invoice details. On Agents' Last Exam it finishes tasks over three times faster than Sol and Astra, and on OSWorld 2.1 OpenAI says Terra beats GPT-5.6 Terra's best score at medium effort. At $1 and $5 per million tokens it also costs half of Sol, and its knowledge cutoff, June 1, 2026, is about a month later than Sol's April 30. Pick Terra for these jobs.

Everyday and high-volume coding

Routine scripted work such as small fixes, test generation and log triage runs at volume, and volume is where the half price shows up. On Terminal-Bench 4.0 at max effort Terra solves 50.3% of tasks, more than double GPT-5.6 Terra's published 21.5%. GitHub's early testing also puts Terra ahead of GPT-5.6 Terra on terminal and C# tasks, but without published numbers, so check it on your own repository before moving a pipeline over. Pick Terra for routine coding and agent steps you can check cheaply.

The hardest coding and reasoning

OpenAI's own advice for the most complex coding tasks points to Sol and Astra. Artificial Analysis puts Sol 4 points higher on its Intelligence Index, 52 against 48, and 6 points higher on its Coding Agent Index, 54 against 48, both at max effort. Sol also costs twice as much per token, so the choice comes down to what a failed run costs you. Pick Sol when a wrong answer costs more than the tokens.

Quick replies with no reasoning

Terra accepts none effort, so it can answer without spending reasoning tokens at all; Sol starts at low. That makes Terra the better fit of the two for classification, short rewrites and chat replies, where a fast answer matters more than depth. Even at low effort, OpenAI's error-inducing prompt set puts Terra's error rate at 16.7%, down from 53.6% for GPT-5.6 Terra. One caveat from Hacker News: in one user's test, Terra spent no reasoning tokens at low either, and low cost less than none (comment). Pick Terra when you want the reasoning step switched off.

Switching between them

• Both run on the same API and endpoints; change only the model id.
• Check reasoning.effort: Sol rejects none.
• A prompt above 272K input tokens is priced at the higher rate for the whole request on both models.
• Batch and Flex halve the price on both.
• In ChatGPT Work and Codex, the same plan allowance covers substantially more tasks with Terra than with Sol, according to the announcement.
• GitHub Copilot offers Terra from October 26, 2026, for Pro, Pro+, Max, Business and Enterprise.
• On OmniaKey, GPT-6.1 Sol is available today; GPT-6.1 Terra is not offered yet (checked October 27, 2026).

Common questions

Is a high-effort Terra as good as a low-effort Sol?

OpenAI publishes no comparison that mixes effort levels across the two models, and Artificial Analysis tests both at max. The question from Hacker News readers is the right one to test, though: run your own tasks at two or three effort levels on each model and compare cost per finished task. The comment thread is where the framing came from.

Is GPT-6.1 Terra cheaper than Sol in practice?

Per token, yes: $1 against $2 for input and $5 against $10 for output. Per task it depends on how many tokens the run spends. On Hacker News, one user said that in their own tests at max effort Terra did worse than Sol, cost twice as much and wrote far longer answers than Sol at higher efforts (comment, follow-up). Compare tokens per finished task before you assume the half price carries over.

Which is faster?

OpenAI says Terra completes Agents' Last Exam tasks over three times faster than Sol and Astra. Artificial Analysis measured 90.9 output tokens per second for Terra against 69.4 for Sol, both at max effort.

What came out first?

GPT-6.1 Sol shipped on September 29, 2026; GPT-6.1 Terra on October 26, 2026. Both sit in the GPT-6.1 family, with GPT-6 Astra above them and GPT-6 Luna below.

For the launch details, read GPT-6.1 Terra: What's New, Benchmarks and How It Compares to Sol.

Update log

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