OpenAI ships GPT-6.1 Terra at $1 per million input tokens, claiming about 9x lower cost per task than Claude Opus 5.5 at max reasoning effort

The source. The primary material is OpenAI's own model page Introducing GPT-6.1 Terra, published on openai.com on October 26, 2026, with the model id gpt-6.1-terra. Every price, percentage, benchmark name and score below is taken from that page; no figures are borrowed from secondary coverage, and nothing the page does not state has been filled in. Independent third-party evaluations, regional availability and enterprise contract terms are not in the page and are therefore absent here.

What it is. The company describes a fast, efficient model that combines detailed visual reasoning, precise computer use and clearer, more natural writing, a significant step up from GPT-5.6 Terra. It is aimed at the tasks that fill a workday: drafting proposals, interpreting technical drawings, filling out forms, editing video. Inside ChatGPT Work and Codex, OpenAI says the same usage allowance covers more tasks than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra. For subscribers that sentence matters more than the API price: same allowance, more work done.

Pricing. Standard API pricing comes in three tiers: $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI states plainly that this is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. The line worth isolating is the cached input price, at 5% of the standard input rate. For work that repeatedly feeds the same long material, checking the same contract clause by clause, asking successive questions against one codebase, or pushing the same invoice template through thousands of documents, the change in cost structure is far larger than the phrase half price suggests, because cached tokens cost almost nothing.

The computer-use line. Agents' Last Exam measures computer use on professional tasks, including turning architectural drawings into a 3D model and reviewing an animated shot frame by frame to produce a time-coded log for a restoration team. The numbers given are: at medium reasoning effort GPT-6.1 Terra scores 15 percentage points above GPT-5.6 Terra at about one third of the cost; it performs on par with GPT-6 Astra at one tenth of Astra's price per token; and at max effort it is more than three times as token efficient as Claude Opus 5.5 at a quarter of its price per token, which compounds to roughly 9x lower cost per task. On speed, it completes tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol.

Vision and desktop tasks. On MMMU-Pro, a visual reasoning benchmark, the company says Terra outperforms GPT-5.6 Terra at every reasoning setting and nearly matches the most capable frontier models, GPT-6 Astra and GPT-6.1 Sol, at much lower cost. On the OSWorld 2.1 offline set, which tests agents driving desktop applications through multistep tasks such as building a 3D model from a technical drawing, at medium reasoning effort it surpasses GPT-5.6 Terra's best score at a fraction of the cost per task. Read together with the computer-use results, Terra's position is clear: it is not the capability ceiling, it is the tier for high-volume, batchable work where errors are handled by retrying.

Factuality and code. OpenAI says it has invested heavily this year in cutting hallucinations, and it put numbers on it: on its internal set of error-inducing prompts the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, a drop of nearly 37 percentage points, and at max effort it fell a further 18 points to 4.3%, close to the rates of GPT-6 Astra and GPT-6.1 Sol. On the code side, Terminal-Bench 4.0 tests coding agents solving problems from the command line, and at max reasoning effort Terra solves 50.3% of tasks, more than double GPT-5.6 Terra's published 21.5%. The company also says outright that for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.

Science tasks and safety. Terminal-Bench-Science 0.1 tests agents doing research tasks in code, such as tracking cell divisions in a time-lapse video. Here OpenAI says Terra at max reasoning effort surpasses GPT-6.1 Sol at lower cost and comes within five points of Opus 5.5 at less than one fifth the estimated cost. On safety, Terra meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework, and keeps the protections developed for GPT-6 Astra. In plain terms, the company acknowledges the model has crossed its own cyber-capability line, and that the safeguards ship with it rather than after it.

How to read the numbers. First, every benchmark here is chosen, run and reported by OpenAI, with no third-party verification, and the cross-vendor curves deserve extra caution: Claude Opus 5.5 and Claude Sonnet 5.5 appear in the charts, but gaps produced under a competitor's own evaluation conditions are not the same thing as product gaps. Second, the company volunteers that the most complex coding work still belongs to its own Sol and Astra models, and that kind of self-limitation usually reads as more credible than a perfect score. Third, the real substance of the price move is not that capability ceilings have converged, but that the unit cost of long agent tasks has collapsed, which changes what products are viable: a budget that used to buy a dozen attempts now buys a hundred.

What this means if you build tools and pipelines. The immediate winners are high-volume, fixed-format, retry-friendly jobs: invoice field extraction, form filling, turning drawings into structured data, batch video shot annotation. Those used to be gated by a few cents per call and now approach zero once the cache is warm; combined with the claimed threefold speedup, the throughput bottleneck shifts from the model to your own orchestration. Three caveats are worth holding onto. Usage allowances inside ChatGPT Work and API billing are two separate ledgers, so do not mix them. The 50% figure is relative to GPT-5.6 Terra and GPT-6.1 Sol, not to every model on the market. And rate limits, regional availability and data retention policies are not in the page at all, which is precisely the part that decides whether something can go to production.