GPT-6.1 Terra Cuts Token Prices. Your Cost per Finished Task Still Needs Testing.

OpenAI's GPT-6.1 Terra announcement makes a concrete offer: cheaper tokens for a model aimed at visual reasoning, writing and computer-based work. That is worth testing. It is not a reason to hand an agent unrestricted access to your accounting software.

The commercial signal is the price, not the launch-day superlative. OpenAI lists standard API rates of $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. The company says standard prices are 50% below GPT-5.6 Terra and GPT-6.1 Sol. Those are vendor-published comparisons. Developer312 has not independently benchmarked this model.

What OpenAI actually announced

According to the launch announcement, the model is available through the API as gpt-6.1-terra, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. The positioning is practical: interpret technical drawings, draft proposals, fill forms and move information between applications.

OpenAI reports that, at medium reasoning effort, Terra scores 15 percentage points above GPT-5.6 Terra on Agents' Last Exam at about one-third the cost. It also reports faster task completion than GPT-6 Astra and GPT-6.1 Sol. These comparisons involve specific evaluation settings. They do not establish the same speed or cost ratios for an arbitrary production workload.

That distinction matters because an agent's bill contains more than its headline token rate. Images, tool calls, reasoning settings, retries and the length of a computer-use trajectory can change total spend. A model that takes more steps may erase a nominal price advantage. A model that makes fewer expensive mistakes may justify a higher token price.

The factuality improvement is promising, not a warranty

OpenAI says its internal error-inducing-prompt benchmark fell from a 53.6% error rate with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra at low reasoning effort. That is a 36.9-percentage-point decline, not a claim that the model is correct 83.3% of the time on every business task.

At maximum effort, the announcement reports a 4.3% error rate on that evaluation. The prompt distribution is crucial. A deliberately difficult internal benchmark measures a particular failure surface; it is not an audited invoice-processing accuracy guarantee. Readers should resist turning one evaluation into a universal reliability number.

The same caution applies to coding. OpenAI reports a 50.3% Terminal-Bench 4.0 score at maximum reasoning effort, compared with a published 21.5% score for GPT-5.6 Terra. That supports the company's claim of improvement on this benchmark. It does not mean half of your engineering backlog can run unattended, or that accepted solutions meet your security and maintainability requirements.

Do the small piece of arithmetic first

Consider an illustrative job using 100,000 uncached input tokens and 20,000 output tokens. At the listed standard rates, the input costs $0.10 and the output costs $0.10: $0.20 in model token charges. This is a calculation, not an observed workload or a complete service quote.

If all 100,000 input tokens qualify for the quoted cached rate, the input component would instead be $0.005. The output component would remain $0.10, producing $0.105 before other charges. Actual caching eligibility and billing behavior must be checked in the applicable API documentation and account terms.

Now add one failed attempt, a correction pass and a reviewer. The spreadsheet changes. This is why procurement should compare cost per accepted result rather than cost per million tokens alone. A five-cent saving is irrelevant if a mistaken payment requires an hour to unwind.

Computer use changes the risk boundary

Computer-use models act through interfaces designed for people. They may encounter customer documents, websites and messages containing untrusted instructions. Reading those materials is not permission to obey them. Good benchmark performance does not remove the need to separate task instructions from external content.

OpenAI says Terra meets its Critical cybersecurity threshold under its Preparedness Framework and retains protections developed for GPT-6 Astra. Its system-card addendum also describes safety evaluations and warns that some comparison values may reflect later versions of earlier models. A capability classification is not a certification that your deployment is safe.

For an invoice workflow, give the agent a limited account, require approval before payment and log the evidence behind every field it enters. For a coding workflow, isolate execution, restrict secrets and review the diff. The model should inherit the smallest authority the task needs, not the full authority of whoever launched it.

How to make the buying decision

Start with a fixed evaluation set drawn from real work: normal documents, messy scans, ambiguous instructions and cases the agent must refuse or escalate. Define acceptance before running the model. Otherwise a convincing answer will quietly become the success criterion.

Track completion rate, total tokens, tool costs, elapsed time, intervention time and material mistakes. Compare the same task distribution and permission boundaries across models. Keep the reasoning setting explicit. A comparison that changes the model, prompt and tool access simultaneously will not tell you what improved.

OpenRouter's model documentation provides a separate distribution listing, including routing-specific terms and prices. It is evidence of a commercial access path, not independent confirmation of OpenAI's benchmark results. Endpoint, region and provider can affect the quote you actually receive.

The bottom line: Terra's listed price makes a controlled trial sensible. Promote it only when your own accepted-result economics improve. Buy verified work, not a cheaper unit of unfinished reasoning.