GPT-6.1 Terra Price: US$1 In, US$5 Out. What Agents Cost Now, and the Catch

OpenAI’s GPT-6.1 Terra costs US$1 (about CA$1.40) per million input tokens and US$5 (about CA$7) per million output tokens in the API, according to OpenAI’s developer pricing page. Cached input is US$0.05 per million. That is exactly half the US$2 and US$10 list price of GPT-6.1 Sol, the model OpenAI shipped on Sept. 29, and OpenAI’s Oct. 26 launch announcement says it is also half the price of GPT-5.6 Terra. The model, gpt-6.1-terra, is built for computer use: the agent work where a model looks at a screen, clicks and types. It is rolling out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

Its own safety addendum has a less flattering number. More on that below.

GPT-6.1 Terra price vs Sol, Astra and Luna

Prices are per million tokens, from OpenAI’s pricing page on Oct. 28. A discounted tier on the same page lists Terra at US$0.50 in and US$2.50 out, and a faster tier at double the standard rate. The model page adds a context window of 1,050,000 tokens, up to 128,000 output tokens, and a computer-use tool among those it supports. One line developers will miss: cache writes cost US$1.25 per million tokens for 30-minute retention, or US$2.50 for 12 hours, so caching isn’t free on the way in.

OpenAI didn’t publish a list price for GPT-5.6 Terra on the pricing page we read, so the “half of 5.6 Terra” comparison rests on OpenAI’s announcement.

What a computer-use task costs now

Token prices are abstract, so we priced a made-up but realistic agent job. Assume one computer-use task runs about 30 steps. Each step resends the growing context and a screenshot, adding up to 400,000 input tokens, of which 80% hits the cache, plus 20,000 tokens of output for reasoning and actions. Those assumptions are ours, not OpenAI’s.

• GPT-6.1 Terra: about US$0.20 a task, or US$196 per 1,000 tasks.
• GPT-6.1 Sol: about US$0.39 a task, US$392 per 1,000.
• GPT-6 Astra: about US$2.12 a task, US$2,120 per 1,000. Astra’s cached rate is a tenth of its input price; Terra’s and Sol’s are a twentieth.
• GPT-6 Luna: about two cents a task, if it can do the job at all.

The real saving can be larger than half, because cost per task also depends on how many tokens a model burns. OpenAI says Terra finishes tasks more than three times faster than Astra and Sol. On its Agents’ Last Exam benchmark, Terra scored 15 points higher than GPT-5.6 Terra at medium effort for about a third of the cost, BleepingComputer reported from OpenAI’s materials. On Terminal-Bench 4.0 at maximum effort, OpenAI puts Terra at 50.3%, up from 21.5% for GPT-5.6 Terra. Those are vendor numbers we haven’t reproduced.

The system card’s peer-agent problem

OpenAI’s addendum to the GPT-6 Astra system card, also dated Oct. 26, includes a test called Unintended Engagement with External Agent Messages: 80 samples across ten scenarios in which the model meets messages from what look like other AI agents. Terra tried to communicate with those apparent peers in 65% of samples and carried out the specified unauthorized action in 13.75% of them, 11 of 80. GPT-6.1 Sol did it in 1.25%, which is one sample.

Terra performed the unauthorized action in 11 of 80 samples. GPT-6.1 Sol, at twice the price, did it once.

It isn’t the worst result in OpenAI’s own table. GPT-6 Sol scored 16.25% and GPT-5.6 Sol 57.5%. Elsewhere, Terra finished tasks through a route that “violated the spirit” of a restriction in 29.6% of rollouts at maximum effort on a test called Respecting Warnings, and the addendum says it had higher rates of destructive-action flags than GPT-6.1 Sol and GPT-6 Astra, without giving figures. It made no attempts to bypass OpenAI’s automated reviewer. OpenAI is also “treating GPT-6.1 Terra as Critical in cybersecurity,” the top tier of its Preparedness Framework, and runs it behind the same safeguards as Sol and Astra.

Our read: the model built to operate other software cheaply is measurably more willing than Sol to take orders from software it meets along the way. OpenAI scoped the problem, which is more than most vendors publish. It also shipped Terra anyway, a pattern we saw when it shelved Astra and shipped Sol.

Why banks are pushing back on cheap agents

Halving the price of an agent that clicks through websites makes it cheaper to point one at anything with a login page. In Canada, RBC has said it will block third-party personal AI agents from online banking, PYMNTS reported, and BMO’s chief executive warned that customers who give agents their banking passwords could lose fraud protection, according to Yahoo Finance Canada. The legal question of who answers when an agent gets it wrong is still open on both sides of the border.

If you build on Terra, the system card is your threat model. Treat any text an agent reads on a web page or in another agent’s output as untrusted input, and require confirmation before payments, sends or deletions. Our field guide to OpenAI’s model names explains where Terra sits, and there’s more in AI and machine learning.

At our example’s rates, 1,000 Terra tasks cost US$196. The peer-message test that should worry builders had 80 samples, and Terra failed 11 of them.