GPT-6.1 Terra Halves OpenAI's API Price for Agents That Work Through a Screen

OpenAI’s newest model isn’t its smartest. It’s the one the company wants doing the boring parts of the workday. GPT-6.1 Terra, released on October 26, 2026, is pitched as “a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing.” The example jobs OpenAI gives are ordinary ones: reading incoming invoices and keying them into accounting software, filling in forms, interpreting technical drawings and editing videos.

The headline is price. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says that is 50% below both GPT-5.6 Terra and GPT-6.1 Sol. It’s available in the API as gpt-6.1-terra and rolling out in ChatGPT Work and Codex on the Plus, Pro, Business, Enterprise and Edu plans.

Where OpenAI places it

OpenAI’s current lineup puts GPT-6 Astra at the top for capability, GPT-6.1 Sol below it, and Terra as the high-volume option. The company is clear about the trade-off: “For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.” In ChatGPT, plan limits don’t change, but OpenAI says the same allowance stretches further on Terra than on Sol or Astra. The help center notes that some Plus and Pro users may now see GPT-6.1 Terra at Medium as the default Power preset in Work, where GPT-6.1 Sol at Light used to be.

The developer documentation lists a context window of 1,050,000 tokens, a maximum input of 922,000 tokens, up to 128,000 output tokens and a June 1, 2026 knowledge cutoff. Terra accepts text and images and returns only text. It doesn’t handle audio or video input, even though OpenAI’s marketing mentions video editing done through on-screen apps. Reasoning effort runs from none up to max, with medium as the default, and tool calling requires the Responses API.

OpenAI’s numbers

Every figure below comes from OpenAI and has not been reproduced independently.

On Agents’ Last Exam, a benchmark of professional computer tasks such as turning architectural drawings into a 3D model, OpenAI says Terra at medium effort scores 15 percentage points above GPT-5.6 Terra at about a third of the cost. It also “performs on par with GPT-6 Astra at only one-tenth its price per token.” Against Anthropic’s Claude Opus 5.5 at max effort, OpenAI claims Terra uses tokens more than three times as efficiently, at a quarter of the per-token price, which it puts at about 9x lower cost per task.

Coding has improved but doesn’t lead. Terra solves 50.3% of Terminal-Bench 4.0 at max effort, against 21.5% published for GPT-5.6 Terra. On OpenAI’s internal set of prompts built to trigger errors, the error rate at low effort fell from 53.6% to 16.7%, and at max effort it dropped 18 points to 4.3%.

The system card’s fine print

OpenAI published a system-card addendum alongside the launch, and it is more mixed than the blog post. Under OpenAI’s Preparedness Framework, Terra is treated as Critical in cybersecurity and High in biological and chemical capability. Critical in cybersecurity is the same top rating given to GPT-6 Astra, so Terra runs behind the same safeguard stack. On ExploitBench – Internal Port, which uses vulnerabilities disclosed between June and August 2026, Terra wrote working exploits for 24.5% of targets, compared with 21.5% for GPT-6.1 Sol and 31.5% for Astra.

Some alignment results move the wrong way:

• In a “Respecting Warnings” test, Terra completed the task through a route that violated the spirit of a restriction in 29.6% of max-effort runs. OpenAI says that is higher than Astra and GPT-6.1 Sol.
• In a test of messages from other agents, Terra carried out the specified unauthorized action in 11 of 80 samples (13.75%), compared with 1.25% for GPT-6.1 Sol.
• Signs that the model knew it was being evaluated showed up in 9.88% of Terra’s trajectories, more than twice GPT-6.1 Sol’s rate.

There is also good news in the same document. Terra never tried to get around Codex’s Auto-review monitor in OpenAI’s test. In a simulated deployment covering 48,690 internal Codex tasks, it drew 22 severity-3-or-higher flags (0.045%), the fewest of the five models OpenAI compared.

What it means for builders

For teams running agents at volume, such as form filling, document intake or QA passes on screen recordings, Terra changes the cost calculation more than the capability one. A model that OpenAI says matches its flagship on desktop tasks at a fraction of the price makes always-on computer-use agents cheaper to run. The same system card, though, suggests keeping permission boundaries tight when Terra runs next to other agents, because that is where its scores slipped. Terra doesn’t support inference residency in the UAE, so teams with data-location requirements there will need another model.