GPT-6.1 Terra Puts AI Agent Costs Under the Microscope

OpenAI released GPT-6.1 Terra on October 26 with a lower standard API price than its predecessor and a pitch aimed at work that happens across screens, images and documents. For developers building agents that repeat those jobs all day, the interesting question is whether the cheaper model can finish enough of them correctly to lower the cost of the work, not just the bill for each unit of text.

Terra is available to developers as gpt-6.1-terra. Its model documentation lists standard rates of $1 per million input tokens and $5 per million output tokens. Tokens are the pieces of text the model reads or generates. OpenAI suggests uses such as matching product photos to catalog entries, interpreting technical drawings, acting on application screens and revising text. It accepts text and image inputs and returns text. Its listed context window is 1,050,000 tokens, meaning it can take a very large prompt, though using all that space has pricing consequences.

Computer use is the other half of the pitch. Rather than just answering a question about an app, an agent can use a computer-use tool to interact with a screen. Terra supports that tool, along with web search, file search and a hosted shell, through OpenAI’s Responses API. Picture a catalog workflow: the model reads a product image, looks at possible matches and helps update an entry. That’s an illustration of the advertised capability, not a claim that Terra reliably completes that entire workflow unattended.

The Price on the Menu Versus the Price of Dinner

A low token rate is useful. An agent that takes several tries, calls paid tools or sends a long prompt can still run up a bill. OpenAI lists cached input at $0.05 per million tokens, with conditions on cache writes. Tool-specific models, including computer use, can carry per-call fees. For prompts exceeding 272,000 input tokens, its documentation says input and cache rates double and the output rate rises by half for the entire request. The enormous context window is therefore a capacity figure, not an invitation to fill it for the base price.

The more revealing numbers concern completed attempts. On Agents’ Last Exam v1, a test of professional computer tasks, OpenAI’s launch-chart figures put Terra at 57.6% and $0.29 per task at medium reasoning effort. GPT-6 Astra also scores 57.6% at that setting, at $4.82 per task; the previous GPT-5.6 Terra scores 42.6% at $0.84. Those task scores and costs are OpenAI-reported benchmark results, not independently established prices or success rates for a company’s own jobs.

Reasoning effort is a setting that changes how much work a model does before responding. Set it higher and the cost-performance pairing can change. At maximum effort on the same exam, Terra reaches 60.0% at $1.01 per task, while Astra reaches 59.3% at $7.24. A buyer choosing a model by the cheapest input token would miss both the near-equal medium-effort scores and the shift in Terra’s cost when its effort goes up.

Those figures suggest a practical experiment: send a defined set of routine screen-and-document jobs to Terra, check which outputs meet the required standard, and compare the total cost with an alternative model. Count attempts, tool calls and any human checking or retries. If an agent must retry a failed entry or ask a person to resolve a questionable match, its first attempt’s price is only part of the expense. That logic follows from how these workflows and charges fit together; the launch-chart task costs cannot tell a team its own retry rate.

Why the Previous Terra Belongs in the Comparison

OpenAI introduced GPT-5.6 Terra in a limited preview on June 26 and opened it to the public on July 9. It then launched GPT-6 Sol and Luna in September without a new Terra. The October 26 release gives developers a successor to compare with a model already positioned for lower-cost work. The new model’s standard API rates are 50% below GPT-5.6 Terra’s.

The predecessor comparison also goes beyond the invoice. At medium effort on Agents’ Last Exam, the reported score rises from 42.6% for GPT-5.6 Terra to 57.6% for GPT-6.1 Terra, while the reported task cost moves from $0.84 to $0.29. On Terminal-Bench 4.0, which tests agents working on coding tasks in a command-line terminal, maximum-effort scores move from 21.5% to 50.3%. OpenAI also reports an improvement on an internal set of prompts designed to induce factual errors. That internal test measures performance on those prompts; it should not be read as an expected error rate for catalog entries or office work.

Taken together, the predecessor figures make a stronger case for testing a switch than a token discount alone would. They pair lower published charges with better reported outcomes on named tasks. They do not decide whether an existing workflow can skip a review step, or whether an agent can safely act on a wrong answer. A team considering a migration has to set its own acceptance criteria before the cheaper attempts start looking like savings.

Astra Still Has Jobs Terra May Struggle to Do

The Astra comparison changes when the job changes. At maximum effort, Terra scores 74.2% on the OSWorld 2.1 offline computer-use test, versus Astra’s 80.9%. On Terminal-Bench 4.0, Terra’s 50.3% trails Astra’s 58.2%, GPT-6.1 Sol’s 55.9% and Claude Sonnet 5.5’s 70.6%. OpenAI points users doing complex coding or in-depth research toward Sol or Astra. An equal score on one professional-task exam is a reason to investigate Terra for similar tasks; it is no promise that it will handle a difficult code change as well as Astra.

There is a narrower visual comparison, too. Terra reaches 85.5% on MMMU-Pro Vision at maximum effort, against Astra’s 86.7%. That could interest someone whose workload involves reading drawings or product images. But identifying what an image depicts and taking the correct action in an application are different steps. A catalog agent might recognize an item and still choose the wrong record or make an unwanted edit. The sensible test would check the whole sequence, including the final change, rather than stop at the model’s answer about the picture.

The choice between models may therefore be made task by task. A developer could reserve a stronger model for work where Terra’s errors prove expensive, while trying Terra on jobs with clear checks and recoverable mistakes. That is a design option, not a measured saving: routing itself needs rules, and the price of checking and escalating work belongs in the comparison. The same published benchmark that makes $0.29 per attempt attractive leaves a substantial share of tasks unsolved at medium effort.

Terra’s launch puts a useful number on the menu, but the decision happens after the order arrives. For an agent clicking through apps, the bill to watch is the one for an accepted result, with the failed attempts and fixes still on it.