OpenAI drops GPT-6.1 Terra for cheap computer use

Running AI agents that actually click buttons, fill out forms, and navigate desktop software has historically had one massive problem: the bill. Burn thousands of tokens watching an agent fumble through an interface, and manual labor suddenly looks cost-effective again.

OpenAI launched GPT-6.1 Terra to tackle that exact friction point. It is a mid-tier model designed specifically for detailed visual reasoning and precise computer use, sitting right between the featherweight Luna and the heavy-duty Sol.

The real headline is the pricing. OpenAI slashed standard API rates by 50% compared to both GPT-5.6 Terra and GPT-6.1 Sol, pricing Terra at $1 per million input tokens, $0.05 per million cached inputs, and $5 per million outputs.

Speed, screens, and fewer hallucinations

Terra is built to take on the tedious screen work that clogs up the average workday, from drafting specs and reading blueprints to handling video logs.

On the Agents' Last Exam benchmark, which tests professional tasks like turning architectural drawings into 3D models or generating frame-by-frame logs for film restoration, Terra scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort. More importantly, it did so at roughly one-third the task cost.

When cranked up to maximum reasoning effort, OpenAI claims Terra runs more than three times faster than both GPT-6 Astra and GPT-6.1 Sol. In head-to-head task costs, OpenAI reported it cost about one-ninth as much per task as Anthropic’s Claude Opus 5.5, while token efficiency was over three times higher.

Early hands-on tests from the community showed the model smoothly steering Blender to generate a 3D tram model straight from prompt instructions.

OpenAI also tackled model reliability. On internal tests with error-inducing prompts at low reasoning effort, Terra cut its factual error rate from 53.6% down to 16.7%. On the command-line benchmark Terminal-Bench 4.0, its solve rate jumped from 21.5% to 50.3%, though OpenAI notes that Sol and Astra remain the go-to choices for the toughest engineering bugs.

Safety-wise, Terra hit the Critical threshold under OpenAI's Preparedness Framework for cyber capabilities and inherits the guardrails built for Astra. It is available right now for Plus, Pro, Enterprise, and Edu tiers across ChatGPT Work and Codex, alongside the API.

Why it matters

We see GPT-6.1 Terra as a calculated, pragmatic move rather than a moonshot.

For everyday developers and businesses, frontier benchmarks do not matter nearly as much as unit economics. If your workflow involves ingesting thousands of scanned invoices, parsing schematics, or syncing data across legacy software, paying frontier model rates kills the business case. Cutting API prices by half while tripling execution speed makes high-volume agentic automation genuinely viable to deploy today.

Still, OpenAI has an uphill battle when it comes to mindshare. Anthropic has captured developer loyalty with its Sonnet and Opus line; OpenRouter data shows Claude volume surged 3.6x after Opus 5.5 arrived. On public leaderboards like Artificial Analysis, Terra still ranks a few points below Claude Sonnet 5.5 in pure capability.

Terra probably will not produce viral benchmark screenshots, and a cheaper mid-tier model may not instantly dethrone Claude from developer dashboards. But it does not need to. By trading frontier bragging rights for speed, low cost, and dependable desktop execution, OpenAI is making a serious bid to become the silent engine running everyday enterprise grunt work.