OpenAI Halves API Prices With Its New Computer-Use Model

OpenAI released GPT-6.1 Terra on 26 October 2026 at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. The company says those standard rates are 50 per cent below both GPT-5.6 Terra, the model Terra replaces, and GPT-6.1 Sol, which sits above it in the range.

The model is built to read a screen and act on it. OpenAI's launch post describes it as "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing", and names drafting proposals, interpreting technical drawings, filling out forms and editing videos as the work it is meant for. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra.

The numbers OpenAI published

On Terminal-Bench 4.0, a test of coding agents working from the command line, OpenAI says Terra solves 50.3 per cent of tasks at maximum reasoning effort against a published 21.5 per cent for GPT-5.6 Terra. On an internal set of error-inducing prompts, the company says the error rate at low effort fell from 53.6 per cent to 16.7 per cent, and at maximum effort fell by 18 percentage points to 4.3 per cent. On Agents' Last Exam, a benchmark of professional tasks that includes turning architectural drawings into a 3D model, OpenAI says Terra at medium effort scores 15 percentage points higher than GPT-5.6 Terra at roughly a third of the cost. Every one of those figures was produced in OpenAI's own harness.

What the independent scoreboard says

Artificial Analysis, which runs its own evaluations, published results the same day. It scored Terra 48 on its Intelligence Index at maximum reasoning effort, against 42 for GPT-5.6 Terra and four points below GPT-6.1 Sol. On its newly released Coding Agent Index v2.0, Terra also scored 48, six points below Sol's 54 and 17 above GPT-6 Luna's 31. Cost per task across the Intelligence Index fell to $0.40 from $1.40 for the older Terra. The firm measured the hallucination rate on its AA-Omniscience test dropping 50 points, with Terra guessing at 38 per cent of the questions it could not answer, down from 88 per cent, while accuracy rose two points.

The gap between the two leagues is narrower than for other releases this month. When Google shipped Gemini 4 Argon on 13 October, as Parallax Nexus reported, it topped the Vals Index while placing joint fourth on Artificial Analysis's index. Terra lands in the same place on both, partly because OpenAI is not claiming the top of either.

OpenAI says so directly. Its launch post tells developers that "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice", and that Terra is there to make high-volume work practical, with reading incoming invoices and entering their details into accounting software as the worked example.

The cheaper token does not always buy a cheaper job. Artificial Analysis found that Terra's maximum-effort setting sits behind its cost-efficiency frontier rather than on it, and that the model wrote 130 million tokens to complete the Intelligence Index where the median comparable model wrote 81 million. A verbose model at half the price can cost the same per finished task.

On safety, OpenAI says Terra meets the Critical threshold in cybersecurity under its Preparedness Framework and keeps the protections developed for GPT-6 Astra, and that in prompt-injection tests it resisted instructions designed to divert it from the user's task. It published no figures alongside that claim, pointing instead to a system card on a separate deployment safety site.

The honest reading is that this is a pricing release rather than a capability release, and OpenAI has been unusually candid about that. Six measured points at under a third of the cost per task is worth a great deal to anyone running thousands of agent loops a day against invoices and web forms, and close to nothing to anyone running one hard problem a week. What neither scoreboard addresses is how Terra behaves across a long computer-use session outside a harness, where one wrong click compounds into a corrupted spreadsheet rather than a failed benchmark row.