OpenAI GPT-6.1 Terra: Computer-Use Model at $1/$5, Half the Price of Sol

On 26 October 2026 OpenAI launched GPT-6.1 Terra, a mid-tier model built for computer use, visual reasoning and coding, at $1 per million input tokens and $5 per million output tokens. BleepingComputer and The Next Web both report that price, which OpenAI says is half that of GPT-5.6 Terra and of GPT-6.1 Sol. Terra is available in the API as gpt-6.1-terra, and to ChatGPT Plus, Pro, Enterprise and Edu users through ChatGPT Work and Codex. It is the model OpenAI wants you to point at forms, spreadsheets and desktop apps, and it arrived a month after OpenAI pulled the launch of GPT-6.1 Astra over safety concerns.

What happened

OpenAI's announcement describes Terra as a fast, efficient model that combines visual reasoning with precise computer use: reading a screen, then clicking, typing and moving between applications. The examples in the coverage are mundane on purpose: filling forms, logging invoices, interpreting technical drawings and automating repetitive desktop work.

Terra replaces GPT-5.6 Terra and sits in the middle of OpenAI's naming scheme, which by now needs a short guide:

• GPT-6 Astra is the flagship, released on 3 September, per TNW's DevDay report.
• GPT-6.1 Sol was announced at DevDay on 29 September at $2 input, $0.10 cached input and $10 output per million tokens, which TNW says is a fifth of Astra's standard rates.
• GPT-6 Sol and GPT-6 Luna are the conversational models that reached ChatGPT on 7 October, 9to5Mac reported, Sol for paid plans and Luna for Free and Go users from the following day.
• GPT-6.1 Terra is the cheaper, faster worker for computer use.

The missing name is GPT-6.1 Astra. On 28 September OpenAI said it would not release it. Saachi Jain, OpenAI's head of safety systems, said in a statement that it did not meet the bar for "scope and authorization, and how it communicates back to the user about the type of work it's done", Al Jazeera reported.

The numbers

Every performance figure below is from OpenAI's own testing, as reported by BleepingComputer and TNW. None has been independently reproduced yet.

OpenAI also says Terra performs on par with GPT-6 Astra at one tenth of Astra's price per token. TNW ties the 15-point gain to a visual reasoning test while BleepingComputer names Agents' Last Exam, so the exact benchmark behind some claims is unclear until OpenAI's full evaluation tables are independently read.

Why it matters

Computer use has been the expensive end of agent work. Each step sends a screenshot, the model reasons over it, and a long task can run to hundreds of steps. Halving the per-token price and cutting the number of seconds per step changes which tasks are worth automating at all. I wrote about where browser and desktop agents actually stand in Browser Agents in 2026; the blocker there was reliability as much as cost, and the hallucination numbers above are OpenAI's direct answer to that.

The safety framing matters too. According to TNW, Terra reaches the "Critical" cybersecurity level in OpenAI's Preparedness Framework and ships with the same protections as GPT-6 Astra. TNW also reports OpenAI saying Terra is more likely to stay within its assigned task and to disclose its limits. That is exactly the behaviour GPT-6.1 Astra failed on, which is why it matters that Terra shipped and Astra 6.1 did not.

What sceptics say

The strongest objection is that all of this is self-reported, and some of it on benchmarks OpenAI does not fully describe. The error rates here come from OpenAI's internal set of questions designed to trip models up, not a field error rate.

Second, Terra is not OpenAI's best coder and OpenAI says so: it still recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding work. Its 50.3% on Terminal-Bench 4.0 is a big step from 21.5%, but other vendors' self-reported numbers on the same benchmark are higher. Moonshot's Kimi K3.1 Pro model card lists 65.7% for Kimi K3.1 Pro, 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5, with different harnesses, so treat cross-vendor comparisons loosely.

Third, the lineup itself is a cost. Four names, two version numbers and a cancelled release in under eight weeks make it hard to pin a model and keep it. Anyone who built on GPT-5.6 Terra now has a successor to evaluate, and anyone who planned around GPT-6.1 Astra has nothing to evaluate at all.

What it means for engineering and DevOps teams

Computer-use agents are the agent class where sandboxing stops being optional. A model that can click and type in a real desktop can also click "delete" in a real desktop. My take, for what it's worth: the cheaper the model gets, the more of these agents people will run, and the more the isolation layer matters. (Disclosure: I build PandaStack, an open-source microVM platform for running agents and untrusted code, so I have a view here.)

• Run computer use in a disposable environment. A fresh VM or container per task, with no long-lived credentials. AI Agent Sandboxes: What They Are and How to Choose One walks through the options.
• Benchmark cost per completed task. Terra's per-token price is low, but screenshots are tokens too. Measure total spend for a representative task end to end.
• Pin the model ID and watch deprecations. Use gpt-6.1-terra explicitly, never a floating alias, and track OpenAI's deprecation notices.
• Keep the hard tasks on a stronger model. Route by task type: Terra for UI automation and routine code, Sol or Astra (or another vendor) for complex changes.

# Pin the model explicitly and log usage, so cost per task is measurable.
curl -s https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6.1-terra","input":"List the steps to export last month invoices from the billing UI."}' \
  | jq '{model, usage}'

FAQ

How much does GPT-6.1 Terra cost?

$1 per million input tokens and $5 per million output tokens in the API, according to BleepingComputer and TNW. OpenAI says that is half the price of GPT-5.6 Terra and GPT-6.1 Sol.

Why did OpenAI cancel GPT-6.1 Astra?

OpenAI said on 28 September that internal testing found the model did not meet its bar on staying within scope and authorisation, and on reporting accurately what it had done. I have not found published test figures for the cancelled model.

Is GPT-6.1 Terra good for coding?

It is much better than GPT-5.6 Terra on OpenAI's numbers, scoring 50.3% on Terminal-Bench 4.0. OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding tasks.

Related: Browser Agents in 2026: 90.2% on OSWorld-Verified, 20.6% on OSWorld 2.0, and Three Agent Browsers Already Shut Down, AI Agent Sandboxes: What They Are and How to Choose One and How to Cut LLM API Costs: Prompt Caching, Batching, Routing (2026).