OpenAI launches GPT-6.1 Terra, a mid-tier model for computer-use agents at half the price

What shipped

OpenAI released GPT-6.1 Terra on October 26, 2026, a mid-tier model built specifically for computer use: reading screens, clicking, typing, and moving between applications to finish multi-step tasks. According to a dev.to explainer, the story of this release is less about raw capability than cost — the pricing is positioned to make routine agent work economically viable at volume.

Pricing

Dev.to lists the API rates as $1 per million input tokens, $5 per million output tokens, and $0.05 per million cached input tokens. That reportedly puts Terra at half the cost of the GPT-5.6 Terra it replaces, which charged $2 in and $12 out per million tokens, at exactly half of GPT-6.1 Sol's rates, and at roughly a tenth of flagship Astra per token. OpenAI also claims, per the article, that the same plan allowance in ChatGPT Work and Codex now covers more tasks with Terra than with its predecessor.

Claimed performance

Every figure in this section is vendor-reported; the dev.to author explicitly notes that independent benchmarks have not been published yet.

On Agents' Last Exam, a set of professional computer tasks, OpenAI reports a 15-percentage-point improvement over GPT-5.6 Terra at medium effort, at about a third of the cost. The company says Terra matches GPT-6 Astra at a tenth of the per-token price and finishes tasks more than three times faster than Astra and Sol.

On Terminal-Bench 4.0, Terra scores 50.3% of tasks at maximum effort, up from 21.5% for its predecessor. On OpenAI's own error-inducing prompts, the reported error rate fell from 53.6% to 16.7% at low effort, and to 4.3% at maximum effort. Against Claude Opus 5.5, OpenAI claims 3x-plus token efficiency at a quarter of the per-token price, which works out to roughly 9x lower cost per task.

GitHub's early testing, cited by dev.to, reportedly found that Terra solved noticeably more terminal and C# tasks than GPT-5.6 Terra and worked through well-scoped problems in focused, consolidated steps.

Availability

The model ID in the API is gpt-6.1-terra. Terra is available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, on Amazon Bedrock, and — as of launch day — generally available in GitHub Copilot for everyday agentic coding and terminal work. Free and Go plans continue to receive GPT-5.6 Terra, and Terra is not offered in Chat itself; it lives in Work and Codex.

Where it fits

Dev.to describes OpenAI's lineup as four tiers: Luna for lightweight work, Terra as the mid-range workhorse, Sol for strong reasoning, and Astra as the flagship. OpenAI's guidance, as reported, is that Sol and Astra remain the picks for the hardest coding and reasoning problems. Terra's pitch is volume: back-office automation that previously did not make financial sense at higher-tier prices — reading invoices into accounting software, filling long forms, interpreting technical drawings.

The article offers three practical recommendations: treat Terra as the default tier for an agent fleet doing routine screen-and-form work; design cache-aware agents, since $0.05 per million cached input tokens makes repeated workflows with shared context dramatically cheaper; and prototype on Terra now but hold off on architectural commitments until third-party benchmarks land, because the reported gains are self-reported.

Why it matters

If the claims hold up, Terra changes the unit economics of running computer-use agents continuously rather than occasionally. At a tenth of flagship per-token pricing, per-task cost stops being the thing that kills automation projects for routine work — the dev.to explainer frames the release as making agent costs cheap enough to stop thinking about.

Two caveats deserve weight. First, this is a single source, and the dev.to post is a condensed developer explainer rather than primary documentation, so the exact benchmark conditions are not spelled out. Second, every performance figure is vendor-reported, and some are large — a Terminal-Bench jump from 21.5% to 50.3% and an error-rate drop from 53.6% to 16.7% are the kind of numbers that warrant independent verification before anyone re-architects an agent stack around them. For now, the reasonable read is that OpenAI has shipped a genuinely cheaper tier for agent volume, with the capability claims still awaiting outside testing.