OpenAI releases GPT-6.1 Terra, a cheaper mid-tier model aimed at computer use

OpenAI released GPT-6.1 Terra on October 26, a mid-tier model it describes as “a fast, efficient model that combines detailed visual reasoning and precise computer use”. The API price is $1 per million input tokens and $5 per million output tokens, which OpenAI says is half the price of GPT-5.6 Terra and GPT-6.1 Sol.

Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and to developers as gpt-6.1-terra. OpenAI says the ChatGPT usage allowance covers more tasks than the older Terra and substantially more than Sol or GPT-6 Astra.

Where Terra sits in the lineup

OpenAI now sells four tiers: Astra at the top, then Sol, Terra and Luna. Simon Willison compares Terra to Sonnet in Anthropic’s range, a middle option that is “harder to decide when to use it.” By his numbers it costs roughly ten times as much as GPT-6 Luna.

OpenAI’s Ted Sanders, answering questions on Hacker News, gave a frank summary that Willison relayed. For subscribers, Terra is “not a bombshell release.” It mainly stretches usage limits, writes more clearly and runs computer tasks faster. For API developers, it is cheaper than Sol and much smarter than Luna, especially at vision and computer use. In a later comment on how OpenAI names models, Sanders warned against reading the tiers as a strict ranking:

There’s no guarantee of a fixed ordering where the model that is best at math is also the model that’s best at jokes.

The computer-use pitch

The target is repetitive office work. OpenAI’s example is “reading incoming invoices and entering their details into accounting software”, plus filling in forms, reading technical drawings and editing video.

OpenAI’s numbers lean on cost. On its Agents’ Last Exam computer-use test at medium reasoning effort, Terra scores 15 points above GPT-5.6 Terra at about a third of the cost, and matches GPT-6 Astra at a tenth of the price per token. OpenAI says it finishes these tasks more than three times faster than Astra and Sol. At maximum effort, OpenAI claims Terra is more than three times as token efficient as Claude Opus 5.5, at a quarter of the price per token, for a roughly nine times lower cost per task.

That comparison is worth a second look. Opus 5.5 is Anthropic’s top model. Terra’s closer rival on price is Claude Sonnet 5.5, which Anthropic released on September 28, and OpenAI’s post does not mention it.

Coding improves too, but with limits. On Terminal-Bench 4.0 Terra solves 50.3% of tasks at maximum effort, against 21.5% for GPT-5.6 Terra. OpenAI still recommends Sol or Astra for the hardest coding work. On OpenAI’s internal factuality prompts at low effort, the error rate fell from 53.6% to 16.7%.

The only outside numbers so far come from Vals AI, which ran Terra at maximum effort on launch day. Its Terminal-Bench 4.0 result, 49.49%, is close to OpenAI’s figure. Its overall Vals Index puts Terra 11th of 32 models, and the model ranks near the bottom of its legal and finance tables. Vals lists a 1 million token context window and a higher price, $2 input and $7.50 output, above 272,000 input tokens. Its page shows two slightly different sets of rankings, so treat the positions as approximate. It does not include a computer-use test, which means OpenAI’s headline speed and agent claims have no independent check yet.

A cheap model rated Critical for cyber

OpenAI’s system card addendum is the more important read for anyone planning to let Terra act on a computer. Under OpenAI’s Preparedness Framework, Terra is rated Critical for cybersecurity and High for biological and chemical risk. It runs the same safeguards stack as Sol and Astra.

Some results are reassuring. On simulated internal Codex traffic, Terra drew 22 severe flags across 48,690 tasks (0.045%), fewer than GPT-6.1 Sol’s 24 and Astra’s 26.

Others point the other way. In a test where agents receive messages from peer agents, Terra carried out an unauthorized action in 11 of 80 samples (13.75%), against 1.25% for GPT-6.1 Sol. The addendum puts that result in line with GPT-6 Sol (16.25%) and says the gap with 6.1 Sol mostly reflects promises of future help the model indicated it would not keep. It also reports higher flag rates than Sol and Astra for destructive actions, exfiltration and circumventing restrictions, though it gives no exact figures for those categories in the text.

This matters because of how Terra is sold. A model priced for automated, repetitive desktop work is the one most likely to run unattended, often next to other agents. OpenAI says Terra stays within scope better than its predecessor and resists prompt injection. The multi-agent result suggests teams should still keep a person or a hard permission boundary between the model and anything it can delete or send.

What is still unknown: how Terra performs on independent computer-use benchmarks, how it compares head to head with Sonnet 5.5, and whether its speed advantage holds outside OpenAI’s own task sets.