OpenAI halves the price of its midrange model and teaches it to click. GPT-6.1 Terra costs $1 per million tokens

OpenAI has shipped GPT-6.1 Terra to ChatGPT Work, Codex and the API. The company pitches it for computer use, reading drawings and writing, and prices it at half of GPT-6.1 Sol in the API. Teams automating forms and invoices now have to redo the math on what their agents are allowed to cost.

Terra costs half of Sol and doubles its command-line score

OpenAI introduced GPT-6.1 Terra on October 26. 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. Simon Willison worked out that Terra is roughly 10x the price of the cheapest model, GPT-6 Luna.

The headline numbers are about computer use. On Agents’ Last Exam, a benchmark that includes tasks such as turning architectural drawings into a 3D model, OpenAI says Terra at medium reasoning effort scores 15 percentage points above GPT-5.6 Terra at about a third of the cost. It reportedly matches GPT-6 Astra at a tenth of the price per token. On Terminal-Bench 4.0 at maximum effort it solves 50.3% of tasks, against a published 21.5% for its predecessor.

OpenAI also reports fewer factual errors. On its internal set of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% to 16.7%. At maximum effort it is 4.3%. The model is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers in the API as gpt-6.1-terra.

A cheap model makes it worth letting an agent click all day

Simon Willison noticed what OpenAI chose to lead with: visual reasoning and computer use. That makes sense. An agent that keeps filling in forms or copying invoices into accounting software burns a lot of tokens on steps that do not need the smartest model. Here cost per task matters more than a benchmark record. OpenAI itself uses reading incoming invoices as its example.

OpenAI’s Ted Sanders explained on Hacker News where Terra fits. In his view it beats Astra and Sol at technical communication, is similar on vision and computer use but much faster, and is a bit worse at most other things. For subscribers, he said, it mainly helps if they keep hitting usage limits. OpenAI adds that Terra finishes tasks more than three times faster than Astra and Sol.

The practical advice for developers is simple. Keep the hardest coding on Sol or Astra, as OpenAI itself recommends, and rerun the numbers for repetitive office agents on Terra. Willison points to the catch: Terra sits in the middle of the lineup, much like Sonnet does at Anthropic, and the middle model is always the hardest one to know when to pick.

Only OpenAI has measured the comparison with Claude Opus 5.5

Every comparison comes from OpenAI, including the choice of benchmarks. The boldest claim, that Terra at maximum effort is about 9x cheaper per task than Claude Opus 5.5, rests on OpenAI’s own measurement of token usage. Independent tests are not out yet. The 16.7% error rate also comes from an internal prompt set nobody outside has seen.

The safety note deserves attention too. OpenAI writes that Terra meets the Critical threshold in cybersecurity under its Preparedness Framework and keeps the protections built for GPT-6 Astra. A cheap, fast model with that rating interests attackers as much as accountants. Only real traffic will show how those protections hold up under a flood of inexpensive queries.

Cost per finished form will matter more than the leaderboard

The key question is whether independent measurements confirm the cost of a whole task, not just the price per token. A model that needs extra attempts can easily burn through a low sticker price. It is also worth watching whether companies move operational agents that run on Sol today over to Terra.

The second signal will come from production. If early reports show Terra quickly and cheaply getting details wrong in forms or invoices, its speed will start working against it.

Lilith's verdict

OpenAI has hired a fast temp for the accounting office at half the wage, one who reads drawings and fills in forms. The savings will only show at month end, when someone counts the invoices that had to be redone after him.