Separate token rates from the cost of a usable result when evaluating Terra: reasoning settings, caching, retries and an explicit acceptance rule.
Lower token rates do not tell you what a finished task will cost. When evaluating GPT-6.1 Terra, compare the attempts needed to produce a result your workflow accepts, alongside the rate card. That is a measurement for your own process, not a conclusion supplied by a model ranking.
OpenAI's model reference lists gpt-6.1-terra, a default reasoning effort of medium, and a requirement to use the Responses API for tool calling. Record these settings before comparing models. Changing the model, effort and tools together makes it harder to explain any difference.
The Standard API rates, checked October 28, are $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. These are separate billing components. Do not apply the cache-read rate to an entire request. Cache writes, long inputs, processing modes and tools can affect the calculation. For a real run, check the applicable rates and usage rather than relying on the model name alone.
Use billable output usage from the API response rather than the length of the visible answer. OpenAI’s reasoning guide says reasoning tokens are billed as output too. If the total already includes them, do not add that component a second time.
Consider a strictly hypothetical request with 10,000 uncached input tokens and 2,000 billable output tokens. At those Standard rates, the token charge would be $0.02. This is not a Terra measurement or a complete bill: it excludes tools, cache writes and other adjustments. Two attempts with identical usage would total $0.04. If only the second passes your check, both attempts belong in the cost of that accepted result.
Define acceptance before testing: for example, correct invoice fields with no missing line items, rather than an answer that sounds convincing. Use the same examples for both models. Retain settings, usage, the cost of every attempt and the accepted-result count; measure manual correction time separately. This prevents a cheaper response from concealing more expensive follow-up work.
Simon Willison published visual examples on October 26 at different reasoning levels. That is one author's experience, not validation of your workload. NODVIEW has not made paid Terra calls: the arithmetic illustrates a comparison method and does not promise savings from a migration.