GPT-6.1 Terra Halves Token Price, But Per-Accepted-Task Cost Can Jump 13x
Token price is no longer a reliable basis for model comparison; per-accepted-task cost, which includes retries and human remediation, determines whether automation actually saves money. Developers who pick models by unit price risk paying 13x more for broken output.
GPT-6.1 Terra arrives with API rates of $1 per million input tokens and $5 per million output tokens, half the price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI also publishes benchmark wins: a 15-point jump on Agents' Last Exam at a third of the cost, a 50.3% solve rate on Terminal-Bench 4.0 versus 21.5%, and a factual-error drop from 53.6% to as low as 4.3% depending on reasoning effort.
The unit-price cut is real, but it does not translate directly into cheaper bug fixes. A task often needs multiple attempts, failed attempts are still billed, and a human has to clean up when the model breaks something. The article walks through a five-step ledger: identify the three billing components (input, cached input, output), record attempt counts, set acceptance criteria before running, log manual remediation minutes, then apply the formula (model cost + manual cost) / accepted tasks. A worked example with a 40% pass-rate model at half price yields a per-accepted-task cost of $3.34, while an 80% pass-rate model at double price costs only $0.25.
The conclusion is a task-based selection rule: small, test-covered fixes and high-volume repetitive work suit Terra's low price, but complex cross-module changes or vague requirements need stronger models like GPT-6.1 Sol or GPT-6 Astra, where failures cost far more in human time. The article also recommends fixing the acceptance environment with a local tool like ServBay and designing reproducible experiments with the same prompt, snapshot, and retry limit across models.
The price war is shifting from per-token costs to per-accepted-task economics, which forces engineering teams to factor in retries and human hours instead of just API rates.
Human remediation dwarfs model cost when pass rates are low; in the example, manual cleanup accounts for nearly the entire 13x difference, so improving pass rate matters more than finding a cheaper token price.
Official benchmark numbers come from specific evaluations and do not guarantee savings on your own codebase; the article's recommendation to build a reproducible acceptance test is the only way to make a defensible model choice.
Model selection has become task-dependent: the same model is economically rational for small, test-covered fixes but costly for complex refactors, so a single global 'best model' no longer exists.
The one-tenth and one-ninth claims in discussions often mix token price with task cost; OpenAI itself uses per-task cost for the Claude comparison, indicating they are not claiming bug fixes are one-tenth cheaper.