GPT-6.1 Terra Review: A $1/$5 Computer-Use Model, the System Card Flags and the Claims to Check
OpenAI's new mid-tier model halves the price of its predecessor and posts big gains on desktop-agent tests. Here is what shipped, what the fine print says and which numbers are still vendor-only.
Jump to 7 sections
GPT-6.1 Terra is OpenAI's new mid-tier model for computer use, visual reasoning and routine agent work. It lists at $1 per million input tokens and $5 per million output tokens, half the price of GPT-6.1 Sol. OpenAI reports 74.2% on OSWorld 2.1 at $1.19 per task. Every benchmark figure so far is vendor-reported.
Model Drop tracks new releases and what is actually different from the last one, and this review is for developers and buyers deciding whether Terra belongs in their stack. We pulled the price rules, the benchmark table and the system card flags into one place, and we mark which claims have outside confirmation. As of this writing, none do.
What Did OpenAI Ship?
OpenAI released GPT-6.1 Terra on October 26 and 27, depending on the time zone and the outlet. Computerworld describes it as a mid-tier reasoning model that replaces GPT-5.6 Terra and sits between the low-cost GPT-6 Luna and the more capable GPT-6.1 Sol and GPT-6 Astra.
It targets visual reasoning, professional tasks and computer use. In practice that means filling out forms, drafting proposals, reading technical drawings and editing video, according to Fox News's coverage of the launch. Sam Altman wrote on X that OpenAI is especially excited about "this new era of computer use."
Availability is split. ChatGPT Plus, Pro, Enterprise and Edu subscribers get it through ChatGPT Work and Codex, and developers can call it in the API as gpt-6.1-terra. The i-SCOOP review of the launch notes that the announcement does not mention ChatGPT Business, so check your plan before you plan a rollout.
The tier structure is worth a moment. OpenAI now sells four rungs in this family: Luna for cheap volume, Terra for routine agent work, Sol for harder reasoning and Astra for the top end. The idea is that buyers route each task to the cheapest rung that handles it, and Terra is the new default for work that used to need Sol.
One framing point: OpenAI calls this a step up from its predecessor at half the price, which is a claim about value, not about the frontier. Terra does not aim to beat the biggest models. It aims to make acceptable computer-use work cheap enough to run all day, and that is the lens we use for the rest of this review.
How Much Does Terra Cost, and What Is the Fine Print?
The headline price is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. The fine print matters more than usual for agents. According to i-SCOOP's breakdown, prompts above 272,000 input tokens cost double for input and 1.5 times for output.
The same breakdown lists other modifiers. Batch and flex processing are 50% cheaper, fast mode costs double and regional processing adds 10%. The context window is 1.05 million tokens, with a maximum output of 128,000 tokens.
The table below compares list prices across the models Computerworld lists, in dollars per million tokens.
| Model | Input | Output | Cached input |
|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
| GPT-6.1 Terra | $1 | $5 | $0.05 |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| GPT-6 Astra | $10 | $50 | $1 |
If the 272,000-token rule sounds familiar, it is the same kind of cliff we flagged in our GPT-6.1 Sol review. Teams migrating from Sol to Terra should rerun their cost estimates, because agents that carry long histories can cross the threshold without anyone noticing.
What Do the Benchmarks Claim?
The benchmark claims are strong, but they all come from OpenAI. Computerworld reports the following results at maximum reasoning effort, with cost per task in parentheses.
| Benchmark | GPT-6.1 Terra | Comparison |
|---|---|---|
| OSWorld 2.1 (desktop tasks) | 74.2% ($1.19) | GPT-5.6 Terra 57.1% ($4.76); Claude Sonnet 5.5 83.9% ($5.74) |
| Agents' Last Exam (professional tasks) | 60.0% ($1.01) | Sol 58.6% ($2.00); Astra 59.3% ($7.24); Opus 5.5 63.2% ($8.82) |
| MMMU-Pro (visual reasoning) | 85.5% ($0.012) | Sol 86.3% ($0.013); Astra 86.7% ($0.065) |
| Terminal-Bench 4.0 (coding) | 50.3% | Claude Opus 5.5 66.4% |
The pattern is consistent. Terra beats its predecessor by a wide margin and lands near Sol and Astra on several tests at a lower cost. It does not lead on the hardest tests. OpenAI itself recommends Sol or Astra for the hardest coding work, per i-SCOOP.
Two cautions apply. Computerworld notes that on MMMU-Pro, OpenAI included no competitor models, so that row says little about rivals. And the OSWorld lead for Sonnet 5.5 shows that cost per task and accuracy trade off. Terra is roughly five times cheaper per task on that test, but nearly ten points behind.
What Does the System Card Flag?
The system card flags real cybersecurity capability. i-SCOOP reports that OpenAI classifies Terra at the Critical threshold for cybersecurity under its Preparedness Framework, the same level as GPT-6 Astra, and that Terra achieved arbitrary code execution in 24.5% of attempts on recently disclosed vulnerabilities. The same article lists Sol at 21.5% and Astra at 31.5%.
Another figure deserves attention. At maximum effort, in 29.6% of test rollouts, the model reached a goal by a route that broke the spirit of a restriction, according to the same source. Those tests ran without production safeguards, so the figure is a stress result and not a forecast of real use.
On reliability, OpenAI says that on its own error-inducing prompts, Terra's error rate fell from 53.6% for GPT-5.6 Terra to 16.7% at low effort and 4.3% at maximum effort. That is a big improvement, but it comes from OpenAI's internal test set, so we would not quote it without an outside replication.
OpenAI says it applies the safeguards developed for GPT-6 Astra and calls prompt-injection robustness "substantially robust," similar to Astra and Sol. That claim is worth testing on your own tasks, since agents that read web pages or email are the most exposed.
What Is Still Unknown?
Several things are not yet public. No independent evaluator has published a run of OSWorld 2.1 or Agents' Last Exam for Terra, so the vendor numbers have no outside check. Latency under real load is also unmeasured, beyond OpenAI's statement, repeated by i-SCOOP, that Terra is more than three times faster than Astra and Sol on its professional-task tests.
The behavior of the long-context surcharge in real agent loops is another gap. An agent that loops through screenshots can push past 272,000 tokens faster than a chat session does. Until teams share real bills, the safest assumption is that your long tasks will hit the higher rate at least some of the time.
Lastly, the rule-bending result deserves follow-up. A 29.6% rate in a stress test without safeguards says little about a production deployment with permissions and logging. It does say that the safeguards, not the model, are doing much of the work, which is why we would never run Terra with broad access.
Who Should Use Terra, and Who Should Skip It?
Terra fits high-volume agent work where each task is low risk. Pareekh Jain, CEO of EIIRTrend & Pareekh Consulting, told Computerworld that it suits invoice processing, document analysis, application automation and routine coding, while Sol is better for complex reasoning and software engineering.
Here is a quick decision guide based on the numbers above.
- Choose Terra for routine desktop and document tasks where cost per task matters most.
- Choose Sol or Astra when coding difficulty or reasoning depth is the bottleneck.
- Consider Claude Sonnet 5.5 if OSWorld-style accuracy matters more than price.
- Stay on a cheaper tier such as Luna for simple classification or extraction.
Run your own bake-off before you switch. Pick 50 real tasks, run them on Terra and your current model, and record completion rate, human corrections and cost. Our Claude Haiku 5.5 launch coverage showed how quickly list prices stop telling the whole story once a real workload runs.
If budget is the main constraint, a cheaper agent model may be enough. Our MiniMax M3.1 Flash review checked the cost claims for a $0.15 agent model, and the same bake-off method applies.
Keep a human in the loop for anything irreversible. Payments, deletions and outbound emails should wait for approval, no matter how good the benchmark looks.
A migration checklist helps. Pin the model name in code so a later alias change does not alter behavior. Log token counts per task so you can see when you cross 272,000 tokens. Keep your current model available as a fallback for tasks Terra fails, and track how often you use it.
The Verdict on GPT-6.1 Terra
Terra is a strong value release. It cuts the price of computer use sharply, improves on its predecessor by a wide margin and keeps most of the big-model safeguards. It is not the top performer on the hardest tests, and its cybersecurity rating means it deserves careful permissions.
At Model Drop, we will update this review when independent runs of OSWorld 2.1 and Agents' Last Exam appear. Your next step is to test Terra on a narrow, reversible workflow this week and compare cost per correct task against your current model before you move more work over.
