GPT-6.1 Terra vs GPT-6.1 Sol

You have an agent that copies 300 supplier invoices a month from email into an accounting system that has no API. The work is visual and repetitive: find the field, type the value, check it registered, move on. Paying for the strongest model on every click feels wasteful. That is the job GPT-6.1 Terra was built for. The question is whether its lower price survives contact with your real workload, or whether GPT-6.1 Sol ends up cheaper because it gets things right the first time.

Quick Answer

  • Choose Terra for high-volume, visual, well-defined computer-use work: form filling, data entry, moving records between business apps, with a human checking anything that submits, pays, or deletes.
  • Choose Sol when the task needs judgment: multi-step reasoning, coding across a codebase, research, or interface work where one mistake is expensive.
  • Use both with model routing: Terra runs the routine screen steps, and Sol takes over when the task gets complicated or Terra’s output fails a check.
  • Choose neither for simple text tasks with no interface involved. A cheaper text model such as GPT-6 Luna will usually do.

There is no universal winner. The right choice depends on your error rate, not the price sheet.

Newer model (November 2026): GPT-6.2 Sol arrived on November 17, 2026 at GPT-6.1 Sol’s price and now starts new ChatGPT Work chats. Terra did not get a 6.2 version. It remains the cheaper computer-use option, and it is one of four models that support OpenAI’s new discounted background completion windows (30% off for a two-hour window, 40% off for eight hours); see ChatGPT.

At a Glance

GPT-6.1 TerraGPT-6.1 Sol
ReleasedOctober 26, 2026September 29, 2026
Built forVisual reasoning, precise computer use, clear writingNear-Astra coding, computer use, and professional work at lower cost
API model IDgpt-6.1-terragpt-6.1-sol
Input / cached / output (per 1M tokens)$1 / $0.05 / $5$2 / $0.10 / $10
Cache writes (per 1M tokens)$1.25$2.50
Context windowAbout 1.05M tokens (922K max input)About 1.05M tokens (922K max input)
Max output128K tokens128K tokens
Computer useYes, hosted tool in the Responses APIYes, hosted tool in the Responses API
Reasoning effort settingsnone, low, medium, high, xhigh, maxAdjustable reasoning effort
In ChatGPTWork and Codex on paid plans, not ChatWork and Codex on paid plans, not Chat

Prices are from OpenAI’s model pages, checked October 31, 2026. Batch and Flex processing are 50% cheaper than standard for Terra, and a faster mode costs twice standard. Inside ChatGPT Work or Codex you are not billed per token; both models draw on your plan’s allowance.

The Main Difference

Sol is a general-purpose workhorse. OpenAI says it approaches GPT-6 Astra on agentic coding, computer use, and professional work at about a fifth of Astra’s price, and Artificial Analysis puts it about a point below Astra on its Intelligence Index.

Terra is narrower and cheaper. OpenAI built it around reading screens and operating software, and says it completes tasks on its Agents’ Last Exam computer-use benchmark more than three times faster than Sol or Astra.

So the trade is speed and price on interface work versus breadth and reliability on everything else.

Computer Use

This is Terra’s home ground, so it is where the comparison is closest.

TaskBetter fitWhy
Filling a stable web form from structured dataTerraFast, repetitive, easy to verify
Entering invoice details into accounting softwareTerra, with checksHigh volume; verify totals and supplier IDs
Updating records in a CRM through its interfaceTerraRoutine navigation and typing
A workflow that branches on what it findsSolNeeds judgment about what to do next
A site that changed its layout since last weekSolRecovering from surprises takes reasoning
Anything that pays, deletes, or sendsEither, with human approvalThe model choice does not remove the risk

Interfaces are messy for any model. Agents misread fields, click the wrong element, miss that a dropdown did not register, and get confused by pop-ups and slow-loading pages. Keep confirmation steps manual and start with dummy data. Our guide to computer-use agents covers the risks in more detail.

Reasoning, Coding, and Research

Sol is the stronger choice here. OpenAI positions it close to its frontier model on agentic coding and professional work, and it was the default model in Codex and ChatGPT Work until GPT-6.2 Sol replaced it on November 17, 2026. Terra can handle simpler steps, but OpenAI does not pitch it as a reasoning or coding model, and the limited independent data (including one accounting benchmark that put Terra slightly behind Sol) points the same way.

If your agent spends most of its time planning, analyzing, or writing code, and only occasionally clicks through an interface, Terra’s speed advantage applies to a small part of the job.

Speed: Task Time, Not Tokens per Second

OpenAI’s “three times faster” claim is about completing computer-use tasks, which is the right way to measure. Tokens per second tells you how fast text streams out; it does not tell you how long it takes to finish filling a form, which also depends on how many steps the model takes, how long it thinks before each action, and whether it has to retry.

Two cautions:

  • The figure comes from OpenAI’s own benchmark. Your interfaces, network, and pages will differ.
  • Terra has a “max” reasoning setting. Turning effort up can make any model slower and more token-hungry, so compare at the settings you would actually use.

Cost: Price per Token vs Cost per Successful Task

On paper, Terra is half the price. In practice, three things can erase that gap:

  1. Token use. Terra can produce more tokens than Sol at high effort settings. If it uses twice the tokens, the saving disappears.
  2. Retries and tool calls. Every failed action that has to be repeated is paid for again, including the screenshots sent back to the model.
  3. Human correction. A person fixing a wrong invoice costs far more than the tokens involved.

The number that matters is cost per completed task:

(total API cost + cost of human corrections) ÷ number of runs that finished correctly

A model that is half the price per token but succeeds 70% of the time can cost more per good result than one that succeeds 95% of the time. Our guide on why you need more than one AI model walks through the arithmetic.

How to Test Terra vs Sol on Your Own Tasks

Public benchmarks will not tell you how either model handles your forms. This experiment will, in an afternoon. We have not published results from it; the point is for you to run it on your own workload.

  1. Build a dummy form with about 15 fields, similar to your real one: text, dates, dropdowns, a checkbox, and one field that only appears after another is filled.
  2. Create synthetic data. Use made-up names, addresses, and amounts. Never use real personal, customer, or financial data in a test.
  3. Write down the correct result for each field before you start.
  4. Run identical instructions with Terra and with Sol, at the same reasoning effort, through the same computer-use setup.
  5. Repeat each run at least 10 times per model, ideally 20, because single runs vary a lot.
  6. Record completion time for each run, from start to the model reporting it is done.
  7. Count correct and incorrect fields in each run, and whether the run as a whole was fully correct.
  8. Record tokens and cost from the API usage data.
  9. Record human corrections: how many fields someone had to fix, and roughly how long it took.
  10. Calculate cost per successful run for each model, including correction time at a realistic hourly rate.

A prompt you can reuse:

You are filling in a web form in the browser window you control.

Goal: Enter the record below into the form at [test form URL] and stop at the review screen.

Record:
- Supplier name: [synthetic value]
- Invoice number: [synthetic value]
- Invoice date: [YYYY-MM-DD]
- Due date: [YYYY-MM-DD]
- Amount (EUR): [synthetic value]
- VAT rate: [synthetic value]
- Cost center: [choose from dropdown: synthetic value]
- Notes: [synthetic value]
(continue for all 15 fields)

Rules:
- Fill every field exactly as given. Do not guess missing values; report them instead.
- After each field, confirm the value was accepted before moving on.
- If a dropdown option does not exist, stop and report it.
- Do NOT click Submit. Stop on the review screen and list every field with the value now shown.
- Finish with: DONE, plus the number of fields you are unsure about.

Stopping at the review screen keeps the test safe and gives you a clean list to compare against your answer key.

Reliability and Where Each One Fails

Both models share the general weaknesses of computer-use agents: misread fields, wrong clicks, broken flows after a website redesign, and the occasional action nobody asked for. The difference is how often, and how well they recover.

  • Terra is most likely to save money on stable, predictable screens where errors are easy to catch automatically, for example by checking totals.
  • Sol is more likely to be worth its price when screens change, data is ambiguous, or one wrong entry triggers expensive cleanup.

A practical pattern is evaluation-gated routing: let Terra do the work, run an automatic check (do the totals match, are required fields filled), and send failed runs to Sol or to a person. That is a form of model escalation.

Who Should Choose Which

Choose GPT-6.1 Terra if you run many similar interface tasks, can verify results automatically, and want lower cost and faster runs. Back-office data entry and form work are the obvious fits.

Choose GPT-6.1 Sol if your agent reasons, codes, researches, or deals with changing or ambiguous interfaces, or if a single mistake costs more than a month of token savings.

Consider a cheaper model if there is no interface involved. For classification, extraction from text, or short drafting, a model like GPT-6 Luna, or an open model such as DeepSeek V4.1 Flash, is cheaper still.

Choose neither, yet, if the task submits payments, deletes data, or contacts customers without a human approval step. Fix the process before picking the model. See AI agent permissions.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See more AI tool comparisons

Browse all side-by-side AI tool comparisons on Ainanza.

Frequently Asked Questions

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

Per token, yes: Terra is $1 per million input tokens and $5 per million output, half of Sol's $2 and $10. Cached input is $0.05 for Terra and $0.10 for Sol. Whether a task ends up cheaper depends on how many tokens each model uses and how often it has to retry, so compare cost per successful task, not price per token.

Is Terra faster than Sol?

For computer use, OpenAI says Terra completes tasks on its Agents' Last Exam computer-use benchmark more than three times faster than GPT-6.1 Sol or GPT-6 Astra. That is OpenAI's own measurement of task completion time. Independent timing data is still limited, and speed on reasoning-heavy or coding tasks may differ.

Is Terra as accurate as Sol?

Not across the board. OpenAI positions Sol as close to GPT-6 Astra on coding, computer use, and professional work, while Terra is tuned for visual reasoning and precise computer use. Early independent results are thin; one third-party accounting benchmark scored Terra slightly below Sol. Expect Sol to be the safer choice for complex reasoning, and test Terra on your own interface tasks.

Can I use Terra in ChatGPT?

Not in the regular Chat tab. As of late October 2026, Terra is rolling out in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu plans, and in the API as gpt-6.1-terra. GPT-6.1 Sol is also available in Work, Codex, and the API but not in Chat.

Do both models support computer use?

Yes. Both can use OpenAI's hosted computer-use tool in the Responses API, and both have a context window of about 1.05 million tokens with up to 128,000 output tokens. Terra is the one OpenAI built specifically around computer use.

Last updated: