OpenAI lists Terra at $1 input and $5 output per million tokens. I read the launch and model card: the useful story is where task cost differs from token price, including the long-context threshold.
§ Article contents
OpenAI introduced GPT-6.1 Terra on October 26. Standard API rates are $1 per million input tokens, $5 per million output tokens and $0.05 per million cached input tokens. This is a new model, not a temporary discount on the same old model. My interest is what happens between a cheap request and a finished piece of work.
The launch positions Terra for visual reasoning, computer use and writing. The model card accepts text and images and returns text; audio and video are not supported modalities. Tool calling requires the Responses API. That distinction matters: a product example about editing video does not mean this endpoint accepts video input directly.
OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum reasoning effort, compared with GPT-5.6 Terra's published 21.5%. Those are vendor-reported results, not tests I ran. The announcement still recommends Sol or Astra for the most complex coding work. I would not turn a launch benchmark into a promise about a particular repository.
Above 272K input tokens, the full request is charged at twice the input and cache rates and 1.5 times the output rate. The advertised million-token context is capacity, not a promise that every context length has the same unit cost.
Cache writes are separate: $1.25 per million tokens for 30-minute retention and $2.50 for 12 hours, before long-context adjustments. Batch and Flex are 50% below Standard; Fast is twice Standard; regional processing adds 10% where available. Tool calls can add fees. A budget needs the route and mode, not only the model name.
I would give Terra and the current model the same small set of real tasks, fixed inputs and acceptance checks. Count accepted results, retries, tool charges and elapsed time. Have a human check the output without knowing which model wrote it. This is a proposed test, not a completed evaluation.
For invoice extraction, an accepted result needs the correct fields and source references. For a screen task, it needs proof that the intended state changed. For code, it needs the relevant test and a readable diff. A low token bill on an unfinished task is still a bill.
Terra earns a place in a controlled comparison because its published rates are lower than the two models named in OpenAI's comparison. It has not earned an automatic production switch. I read the primary pages; I did not call the API or measure quality, latency or billing. Keep that line between a promising price and a proven workflow.
Consider a hypothetical Standard request with 300,000 uncached input tokens and 2,000 output tokens, with no cache writes or tools. The input crosses the threshold. Using the published multipliers gives $0.60 input plus $0.015 output: $0.615 in total. Applying only the headline rates would give $0.31. These are calculated examples, not API invoices.
attempt_costs = [0.615, 0.615, 0.040] # hypothetical USD
accepted = [True, False, True]
count = sum(accepted)
cost_per_accepted = sum(attempt_costs) / count if count else NoneExplanatory Python, not executed against the API. Count every attempt in the numerator, including rejected outputs. The denominator contains only results that pass the acceptance check. Zero accepted results has no finite cost per accepted task.