OpenAI GPT-6.1 Terra Explained: Half-Price Computer Use

On 26 October 2026, OpenAI released GPT-6.1 Terra, a new mid-tier model with a deliberately unglamorous job: doing your office work. Not writing poetry, not solving unsolved science — reading invoices, filling in forms, reading technical drawings, drafting proposals and editing video. The headline is the price. At $1 per million input tokens and $5 per million output tokens, OpenAI says Terra costs 50% less than both the GPT-5.6 Terra it replaces and the flagship-tier GPT-6.1 Sol. The honest story is a trade-off, not a triumph: Terra is cheaper per completed task than the smartest models, and it is not the smartest model. For a lot of real work, that trade-off is exactly right.

What OpenAI actually released

Terra sits in the middle of OpenAI's current line-up, which runs from cheapest to most capable as Luna, Terra, Sol and Astra. GPT-6.1 Terra replaces GPT-5.6 Terra in that middle slot. OpenAI describes it as "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing."

"Computer use" is the part worth understanding properly, because it is easy to over-read. It does not mean Terra has a desk and a mouse. It means the model can look at what is on screen — a form, a document, a drawing, an application window — understand it, and then act inside software: clicking, typing, moving between apps and checking the result. The examples OpenAI gives are deliberately boring, and that is the tell. The model is pitched at the repetitive work that eats an afternoon, not at breakthroughs.

Availability matters more than most launch coverage admits. Terra is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and through the API under the model id gpt-6.1-terra. It is not in regular ChatGPT chat. Availability also varies by plan, region and workspace settings as the rollout proceeds.

On the technical side, the API model carries a 1,050,000-token context window with a maximum output of 128,000 tokens, accepts text and image input, produces text, and has a knowledge cutoff of 1 June 2026, according to OpenAI's developer documentation as reported — enough room for the long document piles that office workflows involve.

The price, in a table you can actually use

API pricing is where Terra's case is clearest. All figures below are per million tokens, as published by OpenAI and reported on launch day:

| Model | Input | Cached input | Output | |---|---|---|---| | GPT-6.1 Terra | $1 | $0.05 | $5 | | GPT-6.1 Sol | $2 | — | $10 | | GPT-5.6 Terra (previous) | $2 | — | $12 |

Terra is therefore half the price of Sol on both input and output, and cheaper than its own predecessor — OpenAI's "50% lower" claim is its own summary of that comparison. Cached input at $0.05 per million is the quiet bargain in the table: when an agent re-reads the same documents across many tasks, repeated input gets dramatically cheaper, which is exactly how high-volume office workflows behave.

But per-token price is the wrong number to budget on. What matters is cost per completed task, because a cheaper model that retries more often, or needs more human correction, can end up costing more. That is where OpenAI's benchmark charts come in — and where they need careful reading.

What OpenAI's own benchmarks claim

Every benchmark number in this section comes from OpenAI itself. No independent benchmark had been published at the time of writing, and that caveat should travel with every figure below.

• Agents' Last Exam (professional tasks), medium effort: Terra scores 57.6%, matching GPT-6 Astra at the same effort level, for about $0.29 per task against Astra's $4.82. It also scores 15 percentage points higher than GPT-5.6 Terra at roughly a third of the cost.
• Agents' Last Exam, maximum effort: Terra scores 60.0% at $1.01 per task, against 63.2% at $8.82 for Claude Opus 5.5. Opus still leads the chart; Terra costs roughly a ninth as much per task.
• Terminal-Bench 4.0 (coding tasks), maximum effort: Terra solves 50.3% of tasks, up from 21.5% for GPT-5.6 Terra. Even so, OpenAI still recommends Sol or Astra for the hardest coding work.
• Factual errors: On OpenAI's internal test of error-inducing prompts, the error rate fell from 53.6% with GPT-5.6 Terra to 16.7% at low reasoning effort, and to 4.3% at maximum effort — close to the flagship models, on OpenAI's own test.
• Speed: OpenAI says Terra completes tasks more than three times faster than Astra and Sol, again on its own timing runs.

Read together, the pattern is consistent: Terra closes much of the gap on routine, well-defined work while staying far cheaper, and the gap that remains sits where you would expect — the hardest, most open-ended problems.

How to try it

If you have access, the practical routes are straightforward:

1. In ChatGPT Work: pick Terra from the model list in Work or the desktop app, rather than in regular chat. Start with a single, well-defined workflow — for example, extracting the same five fields from a batch of similar invoices — and compare the output against your existing process before widening it.
2. In Codex: run codex -m gpt-6.1-terra in the Codex CLI (version 0.159.0 or later, as reported). This is the route for developers wiring Terra into coding or terminal workflows.
3. Via the API: call the model as gpt-6.1-terra. If you are building an agent, log cost per completed task from day one, not just tokens used, and set a review step before any action that moves money, sends messages to customers, or changes records.

Whichever route you take, keep a human in the loop at first. The recurring lesson of this year's agent launches — including Anthropic's proactive Duet agent and OpenAI's own Dots agents — is that an agent that works while you are away needs approval gates and an activity log, not just a good benchmark score.

Who should use Terra, and who should not

Use Terra first if: your work is high-volume, repetitive and well-defined — data entry, form processing, document extraction, routine drafting, or triaging a queue of similar items. Small businesses and operations teams are the natural audience: this is the tier where halving the token price actually halves a visible monthly bill.

Stay on Sol, Astra or Opus if: the task is novel, the cost of an error is high, or the problem needs deep reasoning. Paying more per task is cheap insurance when the task itself is worth thousands.

If you are choosing between vendors: Terra's launch is also a pricing signal. It lands in the middle of an open price war between OpenAI and Anthropic, in which list prices keep falling with each generation while the real price for large buyers is negotiated. Do not switch vendors for a headline rate alone. Run the same workload on both, measure accuracy and cost per completed task, and ask what happens to your price at renewal before you move production work.

What remains uncertain

Three honest unknowns should temper any excitement. First, independent benchmarks have not yet caught up with OpenAI's launch-day claims, so the cost-per-task advantage is currently a vendor chart, not a verified fact. Second, real-world computer use is messier than benchmarks: unusual interfaces and edge-case forms will fail more often than a clean test set suggests. Third, prompts above 272,000 input tokens are charged at higher rates under OpenAI's documentation, as reported, so stuffing an entire archive into one call can quietly erase the savings. OpenAI also says Terra meets the "Critical" cybersecurity capability threshold in its Preparedness Framework — a classification of capability, and another reason a model that operates your software unsupervised deserves permission boundaries, not blind trust.

None of that makes Terra a bad launch. It makes it a mid-tier launch, which is what OpenAI says it is. The models that change a company's costs are rarely the smartest ones. They are the ones cheap and reliable enough to run all day on the work nobody wants to do by hand.

Sources: OpenAI's 26 October 2026 launch announcement and developer documentation, as reported by Skimwire, Blockchain.News, Forkast and AIWeekly. All benchmark figures are OpenAI's own unless stated otherwise. The featured image is an original AI-generated illustration, not a photograph.