GPT-6.1 Terra halves the cost of the agent that reads screens and fills in forms
Released October 26, 2026, GPT-6.1 Terra brings GPT-6-class visual reasoning and computer use at $1 / $5 per million tokens, half the price of Sol. For agents that enter forms and sort documents at volume, it is the first model to evaluate.
October 26, 2026. OpenAI releases GPT-6.1 Terra, a fast model in the GPT-6 family that combines visual reasoning and computer use with clearer writing. 50%. Its standard API prices are half those of GPT-5.6 Terra and GPT-6.1 Sol. $1 / $5. The price per million tokens, in and out. The point that matters: OpenAI is slicing its lineup into cost tiers, and Terra is the tier built for agents that spend their day reading screens and filling in forms — not for reasoning records.
A third price point in the GPT-6.1 lineup
The GPT-6 family has grown by accretion of variants: GPT-6 Astra at the top, then GPT-6.1 Sol, then offshoots like GPT-5.6 Terra. With GPT-6.1 Terra, OpenAI introduces a model that does not try to beat Sol on raw power, but to make high-volume agentic work economically viable.
The positioning fits in one sentence: Terra is the “workday” model — the one that drafts proposals, interprets technical drawings, fills out forms, and edits video. OpenAI is explicitly targeting high-volume flows, such as document entry, form agents, and screen automation, where Sol is “more model than the task needs.” In other words, paying for capacity you never use is becoming a choice rather than a given.
The heart of the model: reading the screen and acting
What sets GPT-6.1 Terra apart is the pairing of visual reasoning and computer use. The model can look at a screenshot, understand what it shows — a dashboard, a form, a technical plan — and act on it, whether filling in fields, matching a product photo to a catalog entry, or interpreting a technical drawing. The context window reaches 1.05 million tokens, enough to load a long thread of screenshots without losing the thread of the task.
The reference profile is consistent with that positioning. According to the BenchLM aggregate ranking of October 28, 2026, GPT-6.1 Terra sits at #9 of 215 models with an overall score of 80.4 / 100. It is strongest in multimodal — 91.6 / 100, #4 of 58, with an 85.5% on MMMU-Pro, the best verified score in that category — and in agentic work: 76.4 / 100, #6 of 134, the 96th percentile. That is exactly the profile of a model built to read documents and screens, not to dominate mathematics.
The numbers, and what they are worth
The raw benchmarks tell the same story, with their caveats. On Terminal-Bench 4.0, Terra scores 50.3%, well behind the best verified score of Claude Sonnet 5.5 at 70.6%. On OSWorld 2.1, the real computer-use evaluation, it posts 74.2%, behind Claude Opus 5.5 at 87.2%. In other words: Terra does not beat the pricier models on pure agent tasks; it wins on value for money, accepting a few points of gap against a bill cut in half.
That is the point to keep in mind. OpenAI is not selling Terra as the most capable model, but as the most cost-efficient one for a specific class of tasks. A data-entry agent processing thousands of documents does not need the best Terminal-Bench score; it needs a reliable model at $5 per million output tokens instead of $10. A 76.4 in agentic work at the 96th percentile is more than enough for that flow.
Terra also lands on Bedrock
The release is not limited to OpenAI’s own API. As of October 26, 2026, AWS announced the general availability of GPT-6.1 Terra on Amazon Bedrock, its managed model platform. In practice, a team that already standardizes its agents on Bedrock — with IAM, guardrails, and unified billing — can move its data-entry flows onto Terra without leaving its existing foundation.
That dual distribution fits the recent strategy: OpenAI is multiplying entry points to its models, and the hyperscalers are hosting the GPT-6 lineup so they do not lose their customers’ agent workloads. For the buyer, it is good news: the same model is available at standard pricing from OpenAI and through Bedrock’s managed pipeline, which makes a total-cost-of-ownership estimate easier before committing.
Against Sol: what to pick at volume
The decision comes down to a trade-off between cost and ceiling. If your agents spend their time reading screenshots, extracting fields from documents, and filling in forms, GPT-6.1 Terra is the best value for money: it is strong exactly where those tasks are won — vision and agentic work — and it costs half as much as Sol. If your workload demands deep reasoning, complex code, or maximum instruction fidelity, stay on GPT-6.1 Sol: those extra points cost money, but they prevent expensive downstream rework.
One blind spot remains: Terra has no independent speed measurements yet, and its Terminal-Bench gap against the priciest competition is real. A purchase should therefore be made on a test bench, not on the spec sheet — run a representative batch of documents, measure the error rate and cost per document, then compare Terra and Sol on the same inputs. That is the only way to know whether halving the price actually halves the real cost.
Verdict
If you run a document-entry or extraction agent at volume, evaluate GPT-6.1 Terra this week: at $1 / $5 per million tokens, it can cut the bill of a flow that runs continuously, without sacrificing what makes those tasks succeed — vision and agentic skill. If your agents must reason, code, or follow complex instructions, keep GPT-6.1 Sol and only step down if a test bench justifies it. The lesson of this release goes beyond the model: the agent race is no longer won on raw capability alone, but on the marginal cost of the automated task.