On 26 October 2026 OpenAI released GPT-6.1 Terra, a fast, inexpensive mid-range model. API prices are half those of GPT-5.6 Terra — $1 per million input tokens and $5 per million output. The company sells it not as the strongest model in the lineup, but as the cheapest way to have a model click through interfaces and read documents for hours a day.
Key takeaways
- API pricing: $1 per million input tokens, $5 per million output, $0.05 for cached input — a quarter of the previous rate
- Artificial Analysis: 48 points on the Intelligence Index, between GPT-6.1 Sol (52) and GPT-5.6 Terra (42)
- Terminal-Bench 4.0: 50.3% of tasks solved at maximum reasoning effort, up from 21.5% for GPT-5.6 Terra
- Guessing when the answer is unknown fell from 88% to 38% on the private AA-Omniscience benchmark
- Available from launch day in ChatGPT Work, Codex and the API for Plus, Pro, Enterprise and Edu plans
Let us picture a situation
An insurance company holds 4,000 scanned claim forms. Each must be opened, three fields read off it, the values typed into an internal system, and the window closed. An AI agent driving the desktop can do this, but at every step it must look at a screenshot, decide where to click, and check the result.
Every one of those looks costs tokens. With 4,000 forms and dozens of steps per form, the bill stops depending on how clever the model is and starts depending on what a single look costs.
But here is the problem: the strongest models are both expensive and slow at this kind of work.
That is the gap GPT-6.1 Terra is aimed at.
What exactly did OpenAI release?
GPT-6.1 Terra is a mid-range model. In the OpenAI portfolio it occupies the slot the Sonnet series holds at Anthropic — neither cheapest nor strongest, historically the weakness of such models: it is hard to decide when to reach for them.
The announcement names three strengths: detailed visual reasoning, precise computer control, and a more natural writing style. The target tasks are deliberately mundane — drafting proposals, interpreting technical drawings, filling out forms, editing video.
We're introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing.
From the OpenAI announcement, quoted on Simon Willison's weblog.
The price is the most telling number in this launch. A million input tokens costs $1, output $5. For comparison, GPT-6.1 Sol and GPT-5.6 Terra are close to twice as expensive, while the cheapest GPT-6 Luna remains roughly ten times cheaper. Input served from cache drops to $0.05 per million — a quarter of the earlier rate.
| Model | Input | Output |
|---|---|---|
| GPT-6.1 Terra | $1 | $5 |
| GPT-5.6 Terra | $2 | $12 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.1 Terra — cached input | $0.05 | — |
What do these names mean?
A handful of benchmark names and terms recur in the coverage of this launch and mean nothing on their own.
- Intelligence Index — a composite score from Artificial Analysis collapsing many tests into one number. Closer to an end-of-term grade than a single exam.
- Pareto frontier — the set of models for which price and quality can no longer both be improved. Stepping off that frontier means giving something up.
- Terminal-Bench 4.0 — a test of how many tasks a model finishes while working in a terminal rather than in a chat window.
- OSWorld 2.1 — a desktop control test: the model receives a screenshot and has to complete a task in a windowed operating system.
- AA-Omniscience — a private benchmark measuring how often a model guesses instead of admitting it does not know.
- Instruction hierarchy — the precedence rule between the provider's, the application owner's and the end user's instructions. The equivalent of filesystem permissions.
Together they answer one question: is the model good enough and cheap enough to be left unsupervised on dull, repetitive work.
Do the numbers back the promise?
Cost of a single task
Artificial Analysis puts the model at 48 points on the Intelligence Index, between GPT-6.1 Sol (52) and GPT-5.6 Terra (42). The ranking matters less than what it costs to reach — at maximum reasoning effort a task comes to $0.40 against $1.40 for the predecessor.
Terminal-based coding
The sharpest jump is in terminal-based coding. Terminal-Bench 4.0 rises from 21.5% of tasks solved to 50.3%, which places Terra in the same tier as the markedly more expensive Sol.
Less guessing
The second change concerns guessing. On AA-Omniscience the share of invented answers fell from 88% to 38%. That is not an improvement in knowledge but in behaviour — the model more often admits it does not know.
Computer use
On computer use, OpenAI reports that on the Agents' Last Exam benchmark Terra matches GPT-6 Astra at one-tenth the token price, and that against Claude Opus 5.5 it is more than three times more token-efficient at a quarter of the token price. The company also claims more than three times faster completion. These comparisons come from the vendor and have not been independently confirmed.
Why not simply always use the strongest model?
The usual choice today: a flagship model for hard work, the cheapest model for easy work. The mid-range model falls out of that scheme because it wins neither category.
Computer use breaks that scheme, because cost scales with the number of steps rather than with task difficulty. One click requires a screenshot, an analysis of it, and a decision. A thousand clicks means a thousand such cycles. A flagship model performs each one slightly better, but the quality gap on a single click is small while the price gap is a multiple.
The second cost is time. A desktop agent holds a resource: a session, a virtual machine, an application window. Three times faster completion means three times fewer occupied sessions, regardless of token price.
That is why Ted Sanders of OpenAI described Terra cautiously in a Hacker News comment: better than Astra and Sol at technical communication, comparable on vision and computer use — but much faster — and somewhat weaker on most other things. For the hardest coding work the company still recommends Sol and Astra.
Is it live yet?
Yes, it is a product available from the day of announcement. The model was released on 26 October 2026 to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, as well as through the API.
Terra meets the “Critical” cybersecurity threshold under OpenAI's Preparedness Framework and inherits GPT-6 Astra's safeguards. Per the system card, its defence rate against indirect prompt injection — malicious instructions hidden in third-party content — is 98.8%, up from 96.2% for GPT-5.6 Terra.
Why does it matter?
For two years progress was measured by how high a score went. This launch measures it by the ratio of score to bill, and the target category is not reasoning but clicking. That shifts the question from whether a model can do something to how many times it can be allowed to try — a different economics of deployment.
What comes next?
- OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra for the hardest coding tasks — Terra replaces neither in the price list.
- Artificial Analysis placed the model on the Pareto frontier for lower reasoning-effort levels, which means the next price cut in this class will have to come with either a quality drop or a new architecture.
- Cached input at $0.05 per million tokens makes repetitive work with a long, stable context the cheapest use case — the first area where comparisons with Anthropic will appear.
Sources
- The Decoder — GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger models
- Simon Willison's Weblog — GPT-6.1 Terra
- OpenAI — Introducing GPT-6.1 Terra
