OpenAI has released GPT-6.1 Terra: A faster, more affordable tier for computer-use and batch desktop tasks; the standard API price is $1/M for input, $0.05/M for cache input, $5/M for output(Subject to the information provided on the official website on that day.) To get straight to the point:For batch Agent tasks, don’t just “switch everything over to the cheaper option as soon as you see it”—first use the same evaluation set to check completion rates and failure retry rates, then calculate the cost per task.. This article isComparative Structure: Flagship tier vs. budget computer-use tier—compare completion rates, cost per task, and failure retry rates in a three-way comparison. Do not fabricate percentages for internal incidents; cite only verifiable, publicly available data. The last two sections include 20 evaluations you can complete tonight and the first-week acceptance criteria.

Don't Rush to Make a Full-Scale Switch Just Yet: What the Budget Tier Solves—and What It Doesn't
The appeal of the budget tier is clear: for the same set of tasks—such as “processing invoices,” “filling out forms,” and “navigating interfaces”—the company claims it offers significant advantages in both price and speed compared to the previous-generation Terra and other flagships in the same series. After a demo on the first day, the manager asked, “Can we switch over completely tomorrow?” The problem is thatTask Distribution: Some steps can be completed through visual orientation alone, while others require extensive reasoning, rigorous fact-checking, or complex coding—the latter may still be better suited for flagship or specialized coding tracks.
In its public presentation, Terra emphasizes visual reasoning, computer use, and clearer writing, and provides comparisons based on metrics such as Agents’ Last Exam, OSWorld, and Terminal-Bench. These areManufacturer Reviews and Pricing Narratives...not the completion rate of those 20 business tasks in your warehouse. Treating the official website score as if it were a guarantee that “we’ll go up 15 points tomorrow” is like treating someone else’s exam paper as your own diploma.
Another type of false sense of security is the belief that “since the unit price is lower, the total cost must be lower.” However, if the low-cost option results in a high failure rate, frequent retries, and extensive manual intervention, the unitSuccessful MissionsThe cost might actually be higher. In 3D, the “unit mission cost” must refer to the cost of a successful mission, not the token price.
Things to Do Tonight:List the types of Agent tasks you perform most frequently, and indicate whether they are “primarily computer-use” or “primarily long-form reasoning/coding.” The latter will not be included in the full list of candidates by default.
Three-Dimensional Comparison: Completion Rate, Cost per Task, and Retry Rate After Failure

Lay out the two paths on a table:
| dimension (math.) | Flagship / High-Performance | Affordable computer-use software (such as Terra) |
|---|---|---|
| What Factors Determine the Completion Rate? | Will complex reasoning, rigorous facts, and difficult coding steps be completed on the first try? | Has visual positioning, forms, and repeated desktop streams ever been stable? |
| Unit Task Cost | Tokens are expensive, but they may actually save you money when there are fewer failures. | Low list price; use "Completed Tasks" to apportion costs, including retries |
| Failure Retry | The cost of occasional failures is high, so limits should be set. | When failures are more common, retries can eat into the price spread. |
| My Default Order | Challenging Tasks and the Return of the Gold Standard Still Stand in Contrast to the Flagship | High-throughput, well-structured computer-use early graying |
This comparison isn’t to say that flagship models are useless, nor that budget models are mere toys. Official materials position the Terra on the “cost-intelligence frontier” and on the side of large-scale computer usage; when it comes to “the most complex coding,” they still single out higher-end models. Your decision should beRouting by Task Bucket...rather than a one-size-fits-all approach at the repository level.
Things to Do Tonight:Use the table above to check off the current top 5 task types—whichever column already has evidence. For the column lacking evaluation, write the name of tomorrow’s task instead of “Just get it done first.”
How to Read Publicly Available Materials: Pricing, Positioning, and Evaluation Criteria

When reading publicly available materials, I first focus on three key points, then apply them to my own evaluation set:
- Pricing and Model ID:The official website states that the API price and
gpt-6.1-terraAvailability: As of the day you open the page; include a screenshot in the decision-making document. - Positioning Boundaries:In which scenarios is the official recommendation "strong" (computer use, visual tasks, batch office workflows), and in which scenarios is a higher tier still recommended? The boundary cases are more important than the scores.
- Review Criteria:Agents’ Last Exam, OSWorld, Terminal-Bench, and others are public benchmarks, not business SLAs. You need to come up with your own 20 criteria.
When integrating with Responses / computer_use in your project, follow the instructions in your existing SDK and cookbook examples; do not run three sets of clients in parallel. Include the model name, tool declarations, and retry limit in a single file. agent-routing.mdThe

Things to Do Tonight:Copy the three lines of pricing from the official website and the three sentences each for "Applies" and "Does Not Apply" into agent-routing.md, and indicate the date of the excerpt (Asia/Shanghai).
Coming tonight: Let’s run 20 tests on the same test suite before discussing the switch.

Don’t wait for next week’s “Model Selection Workshop.” Spend 90 minutes tonight following these three steps, and only then will you be ready to discuss a full-scale migration tomorrow.
- Fixed Test Set (25 minutes):Select 20 real-world tasks (anonymization allowed), covering clicking, form filling, image recognition and data entry, and at least 3 “known difficult cases.” For each task, describe the expected observable outcomes; responses that merely “seem acceptable” are not permitted.
- Dual-Model Same Set (40 minutes):Run one round each on the flagship tier and the budget tier for computer use; record success/failure, number of retries, token, or billing estimate. Include the commands in the README for one-click execution locally.
- Calculate the cost of a successful mission (25 minutes):Calculate the average cost only for successful samples; for failed samples that enter the manual queue, add the number of minutes spent on manual processing. Whichever team wins in terms of “meeting the completion rate ∩ lower cost per success” will be added to the gray-list.
# Example: In-suite comparison (command names may vary depending on your repository) python3 scripts/run_agent_eval.py --suite evals/cua20.jsonl --model flagship --out out/flagship.jsonl python3 scripts/run_agent_eval.py --suite evals/cua20.jsonl --model cheap-cua --out out/cheap.jsonl python3 scripts/compare_cost_success.py --a out/flagship.jsonl --b out/cheap.jsonl
Things to Do Tonight:First, run the three "smoking" tests to confirm that the model ID and the number of retries appear in the logs, then run all 20 tests.
Week 1 Acceptance Testing and Troubleshooting
- Routing tables in Git: Which task buckets are set to the "Economy" tier by default, which are required to be "Premium," and what is the gray-hair ratio?;
- If the completion rate threshold is not met, submitting a "full migration" PR is prohibited;
- Run the same 20 prompts every week; any changes to the model or prompts must be accompanied by a before-and-after comparison;
- If you notice an anomaly on your bill, first check for duplicates using "Retry Storm," then check the unit price.
Troubleshooting System Crashes
- Low List Prices, but High Bills:Failure Retry and Excessively Long Trajectories. Set an upper limit for retries; if a timeout occurs, escalate to manual handling.
- High official website score, low business score:The task distribution is different. Provide examples that increase the difficulty, rather than criticizing the assessment.
- Coding Tasks Collapse After Full Cut:The code bucket should be moved back to the Flagship or Codex classpath; do not leave it in the "computer-use" folder.
- The comparison script cannot be reproduced:Missing fixed seed and model ID logs. Fix the logs first, then discuss the decision.
Final adjustments to the sequence: Find a cheaper option → 20 items in the same category → 3D comparison → Sort by bucket of gray hair → Expand once the target is met. Terra has driven down the price for "computer-use" items; what’s missing is the review you’re starting tonight. We can write a more polished narrative transition later; for now, let’s make the failed retries visible in the table.


Comment List (0):
Load More Comments Loading...