On 26 October 2026 introduced Terra, a model built for office work on a computer: filling out forms, reading documents, interpreting screens and writing. In the API it costs half as much as . We worked out three typical small-business tasks to see what it now costs to have an agent work on your website or in your back office, and compared it with ’s models.
What OpenAI announced
According to OpenAI’s announcement, Terra combines visual reasoning, computer use and more natural writing. As an example of high-volume work, mentions reading incoming invoices and entering their details into accounting software. It is available in Work and on the Plus, Pro, Enterprise and Edu plans, and in the API as gpt-6.1-terra.
The quality figures below are ’s, measured in its own tests:
- Fewer errors. On its internal benchmark of error-inducing prompts, the error rate falls from Terra’s 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum effort.
- Computer use. On the Agents’ Last Exam benchmark, at medium reasoning effort, it scores 15 points higher than Terra at about a third of the cost. At maximum effort and against , estimates a cost per task about 9 times lower.
- Malicious instructions. says that in its prompt-injection tests (hidden text on a website or in a document that tries to divert the agent) Terra showed strong resistance. The announcement gives no figure.
Official prices per million tokens
This is what the OpenAI and Anthropic pricing pages say today, in US dollars and in standard mode, for normal requests (up to 272,000 input on ):
- Terra: $1 input, $0.05 cached input and $5 output.
- : $2, $0.10 and $10. Terra costs exactly half.
- : $4, $0.20 and $20.
- : $2, $0.10 and $10, the same as Sol.
- : $0.10, $0.01 and $0.50 (for prompts up to 100,000 tokens).
says Terra costs 50% less than Terra. In the table that holds for input ($2 against $1); on output the cut is bigger, because Terra charges $12.
Three small-business tasks, with numbers
The calculations are ours, not ’s or ’s. We assume a reasonable number of tokens for each task and apply the prices above. Reasoning tokens are billed as output, so we include them there.
- Answering a customer in your website chatbot. 2,000 tokens of instructions and documents repeated with every question (cached), 1,000 tokens of conversation and 400 of answer.
- Reading an invoice and preparing its details for accounting. 6,000 input tokens (the scanned invoice and the instructions) and 1,000 output.
- Filling out a web form with a customer’s details. An agent that looks at the screen step by step: 30,000 input tokens, another 10,000 cached and 3,000 output.
For every 1,000 tasks, the model costs:
- Chatbot: $3.10 with Terra, $6.20 with Sol or Sonnet 5.5, $12.40 with Opus 5.5 and $0.32 with Haiku 5.5.
- Invoice: $11 with Terra, $22 with Sol or Sonnet 5.5, $44 with Opus 5.5 and $1.10 with Haiku 5.5.
- Form: $45.50 with Terra, $91 with Sol or Sonnet 5.5, $182 with Opus 5.5 and $4.60 with Haiku 5.5.
For a small business: a shop with 1,500 chatbot replies a month would pay about $4.65 in model costs with Terra. An accounting firm processing 400 invoices a month, about $4.40. And 200 forms a month, about $9.10.
What the table does not tell you
Price per token is not cost per task. Four things move the sum:
- How many tokens each model uses. A model that reasons more, or makes mistakes and repeats steps, spends more even if its price is lower. also warns that its models from onwards use a tokenizer that produces around 30% more tokens for the same text.
- Whether the result is usable. Haiku 5.5 is ten times cheaper than Terra, but if it makes more mistakes on your form, the savings go on reviews. Only testing with your own cases tells you.
- Discounts and surcharges. At both companies, batch processing (for tasks that are not urgent, such as the day’s invoices) costs half. charges 10% more if you require data to be processed in a specific region, and applies 1.1 times if you pin inference to the United States.
- Everything else. Computer-use tools add fixed tokens to every request, and the model is only part of the spend: integration, maintenance and review weigh more. In what an AI agent really costs we break it down.
What to do with this
If you already run a chatbot or an automated process on a more expensive model, it is worth testing Terra with your own cases and comparing results, not just prices. If you are starting from scratch, the model cost is rarely the obstacle any more: having your data in order and deciding what the agent does when it does not know something are. And with agents that read third-party websites and documents, check their permissions: resistance to malicious instructions is improving, but no vendor claims it is complete.
Where we fit in
If you already have a chatbot on your website and you want to fix something specific, such as the model it uses, its cost or its answers, our One-off Review (€150) fits: it is meant for targeted fixes with a single delivery. If you want a full report on your website or app, there is the Technical Audit (€200). And if you are looking for a new chatbot for your website connected to your data, see how we build chatbots or get in touch.
