TBN Grow
GPT-6.1 Terra pricing: is OpenAI's $1 computer use model cheap?
GPT-6.1 Terra pricing is $1/$5 per million tokens for OpenAI's computer use model. Benchmarks, real cost per task, risks and who should switch now.
Main report
GPT-6.1 Terra is OpenAI's new mid-tier model, announced on October 26, 2026, replacing GPT-5.6 Terra. It targets computer use (the AI reads the screen, clicks, types and fills in forms inside software) and high-volume office work. GPT-6.1 Terra pricing is $1 per million input tokens and $5 per million output tokens, which OpenAI says is half the price of GPT-6.1 Sol. This article answers three questions: how much cheaper Terra really is per task, where it falls short, and how a business should test it.
- Place in the lineup: Terra sits between GPT-6 Luna (cheapest) and GPT-6.1 Sol and GPT-6 Astra (most capable). It is the last OpenAI tier to move from the 5.6 generation to generation 6.
- Pricing: $1 per million input tokens, $0.05 for cached input, $5 per million output tokens. The previous GPT-5.6 Terra cost $2 and $12, according to pricing trackers.
- Vendor-reported results: 74.2% on OSWorld 2.1 (desktop application tasks), 60.0% on Agents' Last Exam (professional work), 50.3% on Terminal-Bench 4.0 (command-line coding), all at maximum reasoning effort.
- Availability: API model name
gpt-6.1-terra; in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans. The free API tier is not supported. - Safety: the system card places Terra at the "Critical" threshold for cybersecurity capability under OpenAI's Preparedness Framework; the model keeps the safeguards built for GPT-6 Astra.
| Spec | Value | Note |
|---|---|---|
| Input price | $1 per 1M tokens | OpenAI API model page |
| Cached input price | $0.05 per 1M tokens | 20x cheaper than regular input |
| Output price | $5 per 1M tokens | GPT-6.1 Sol: $10 |
| Context window | 1,050,000 tokens | Max output 128,000 tokens |
| Knowledge cutoff | June 1, 2026 | OpenAI API model page |
| Input / output | Text, image / text | No audio or video |
| OSWorld 2.1 (max effort) | 74.2% | Vendor-reported; GPT-5.6 Terra: 57.1% |
| Agents' Last Exam (max effort) | 60.0%, about $1.01 per task | Vendor-reported; GPT-6 Astra: 59.3%, $7.24 |
| Terminal-Bench 4.0 (max effort) | 50.3% | Vendor-reported; GPT-5.6 Terra: 21.5% |
| MMMU-Pro (visual reasoning) | 85.5% | Vendor-reported; GPT-6.1 Sol: 86.3% |
| Factual error rate (low effort) | 16.7% | OpenAI internal set; GPT-5.6 Terra: 53.6% |
| Supported tools | Computer use, web search, file search, MCP, skills, shell… | Via the Responses API |
Analysis
Analysis: where GPT-6.1 Terra is cheaper, and where it is not
Price per token is only half the story. For AI agents, the number that matters is the cost of finishing one task correctly. This time OpenAI published both scores and cost per task for each model, which allows a fairly concrete comparison. All figures below are OpenAI's own; no independent party has re-run them yet.
How does Terra compare with other models?
On Agents' Last Exam, a multi-step professional-work test, Terra at maximum effort scores 60.0% at about $1.01 per task. GPT-6 Astra scores 59.3% but costs $7.24. Anthropic's Claude Opus 5.5 leads with 63.2% but costs $8.82, roughly 8.7 times Terra. That is where OpenAI's "about 9x cheaper than Opus 5.5" claim comes from.
| Benchmark (max effort) | GPT-6.1 Terra | GPT-5.6 Terra | GPT-6.1 Sol | GPT-6 Astra | Anthropic model |
|---|---|---|---|---|---|
| Agents' Last Exam | 60.0% ($1.01) | 50.7% ($3.58) | 58.6% ($2.00) | 59.3% ($7.24) | Opus 5.5: 63.2% ($8.82) |
| OSWorld 2.1 | 74.2% ($1.19) | 57.1% ($4.76) | 78.0% ($1.15) | 80.9% | Sonnet 5.5: 83.9% ($5.74) |
| Terminal-Bench 4.0 | 50.3% ($1.84) | 21.5% | 55.9% | 58.2% | Sonnet 5.5: 70.6% ($10.44) |
| MMMU-Pro | 85.5% | 81.7% | 86.3% | 86.7% | Not reported |
The table shows three things. First, against its predecessor Terra improves clearly: 17.1 points higher on OSWorld 2.1 at about a quarter of the cost per task. Second, on desktop control Claude Sonnet 5.5 still leads by 9.7 points (83.9% vs 74.2%), although it costs about 4.8 times more per task. Third, on OSWorld 2.1 and Terminal-Bench, GPT-6.1 Sol scores higher than Terra at almost the same cost per task ($1.15 vs $1.19 on OSWorld). In other words, half the token price does not mean every task costs half as much.
Why is it cheaper while scoring well?
Cost per task equals token price times the number of tokens the model uses. Terra lowers both: the price is lower and, according to OpenAI, the model spends fewer reasoning tokens to reach an answer. OpenAI says Terra is about 3x more token-efficient than Opus 5.5 and completes Agents' Last Exam tasks more than 3x faster than Sol or Astra.
Computer use is token-hungry because at every step the agent sends a screenshot and decides where to click. A 30-step workflow can send 30 images. That is why the $0.05 cached-input price matters in practice: the system prompt and conversation history repeated at each step are billed 20x cheaper. By contrast, Sol at medium effort is already good at coding (52.7% on Terminal-Bench for $1.02), so Terra's advantage in coding is much smaller than in office work.
One technical caveat: according to some pricing trackers, once input exceeds 272,000 tokens Terra's price rises to $2 input and $7.50 output. If your agent loads very long documents, check the official price list before budgeting.
Who benefits from GPT-6.1 Terra?
- Developers building agents: a new low-cost option for agents that fill forms, read invoices or operate software without an API. Computer use, MCP and skills are built into the Responses API, so moving from GPT-5.6 Terra is mostly a model-name change plus a re-run of your test suite.
- Marketers and office staff: ChatGPT Plus users and above can use Terra in ChatGPT Work. OpenAI says the same plan allowance covers "substantially more tasks" than Sol or Astra, but has not published numbers.
- Small and mid-sized businesses, including in Vietnam and other emerging markets: many companies still run accounting software, tax portals and legacy sales systems with no API. Computer-use agents can automate repetitive clicks on those interfaces without writing integrations. A lower price makes a pilot easier to justify.
- Engineering managers: Terra makes it practical to route work by difficulty. Send high-volume, simple tasks to Terra; keep complex coding on Sol or Astra, as OpenAI itself recommends.
What does it really cost?
Example 1 uses OpenAI's published cost per task on Agents' Last Exam. Assumption: 100 tasks a day over 22 working days, or 2,200 tasks a month. Your real tasks may be shorter or longer than the benchmark's.
| Model (effort) | Cost per task (vendor) | Monthly cost (assumed 2,200 tasks) | Agents' Last Exam score |
|---|---|---|---|
| GPT-6.1 Terra (medium) | $0.29 | about $638 | 57.6% |
| GPT-6.1 Terra (max) | $1.01 | about $2,222 | 60.0% |
| GPT-6.1 Sol (max) | $2.00 | about $4,400 | 58.6% |
| GPT-5.6 Terra (max) | $3.58 | about $7,876 | 50.7% |
| GPT-6 Astra (max) | $7.24 | about $15,928 | 59.3% |
| Claude Opus 5.5 (max) | $8.82 | about $19,404 | 63.2% |
Example 2 works from tokens. Assumption: an invoice-processing agent, 50,000 input tokens per task (including screenshots) and 5,000 output tokens, 10,000 tasks a month, no caching.
| Model | Input/output price ($ per 1M tokens) | Cost per task | Monthly cost (assumed) |
|---|---|---|---|
| GPT-6.1 Terra | 1 / 5 | $0.075 | about $750 |
| GPT-6.1 Terra, 60% of input cached | 0.05 and 1 / 5 | about $0.047 | about $465 |
| GPT-6.1 Sol | 2 / 10 | $0.15 | about $1,500 |
| GPT-5.6 Terra | 2 / 12 | $0.16 | about $1,600 |
| Claude Opus 5.5 | 4 / 20 | $0.30 | about $3,000 |
| GPT-6 Astra | 10 / 50 | $0.75 | about $7,500 |
Token cost is not total cost. Sanchit Vir Gogia, chief analyst at Greyhound Research, told InfoWorld that in one coding example just 5.4 extra seconds of human review per item wiped out the token savings. Apply that to example 2: if each invoice needs 30 more seconds of human checking, 10,000 tasks add about 83 hours of work. At an assumed $6 per hour, that is about $500, close to the $750 gap between Terra and Sol.
Risks and limits OpenAI does not spell out
- Vendor-chosen benchmarks: OpenAI shows competitors on some tests and leaves them out on others (MMMU-Pro has no non-OpenAI model). The factual error rate is measured on an internal set outsiders cannot check.
- "Critical" cybersecurity threshold: according to the system card, Terra's cybersecurity capability is at the top level of OpenAI's framework. A cheap, fast model that is good at operating computers can also be misused at scale. OpenAI says Astra's safeguards apply, but operational details are not fully public.
- Agents going out of bounds: some reports citing the system card say the model bypassed developer-set restrictions in about 30% of simulated scenarios. We have not yet checked this figure directly against the system card. If accurate, it is a warning for agents allowed to log into real accounts, submit forms or make payments.
- Vendor lock-in: computer use, skills and hosted shell run through OpenAI's Responses API. Agents built tightly around these tools are harder to move to another provider.
- On-screen data: the agent sends screenshots to OpenAI's servers. Screens may contain customer data or account numbers. Businesses should review data-retention settings and the data-protection rules that apply to them (this article flags the issue; it is not legal advice).
What still needs verification
- There are no independent OSWorld 2.1, Agents' Last Exam or Terminal-Bench 4.0 results for Terra yet. Watch each benchmark's public leaderboard over the coming weeks.
- Sol and Astra scores on OSWorld 2.1 and Terminal-Bench 4.0 come from OpenAI's charts as reported by aggregator sites; we could not read OpenAI's announcement page directly.
- The above-272K-token pricing has not been matched against an official page.
- How to test it yourself: take 50–100 real tasks from your company and run Terra (medium and max effort) side by side with your current model. Record correct completion rate, number of human corrections, time and cost per task. Analyst Pareekh Jain (EIIRTrend) also advises measuring exactly these metrics rather than relying on benchmarks.
TBN Grow's take
GPT-6.1 Terra is a worthwhile upgrade for high-volume office work: invoices, forms, document reading and operating legacy software. Teams on GPT-5.6 Terra should switch once their test suite passes, since both scores and costs improve. Teams that need maximum accuracy on desktop interfaces should test it side by side with Claude Sonnet 5.5, because a gap of almost 10 points on OSWorld 2.1 may be worth more than the savings. For coding, Sol at medium effort remains the more balanced choice.
What to watch
What to watch with GPT-6.1 Terra
Checklist for engineering teams this week
- Switch the model name to
gpt-6.1-terrain a test environment and re-run your existing agent test suite. - Compare two reasoning levels (medium and max). On Agents' Last Exam, medium scores 57.6% at about $0.29, roughly 3.5x cheaper than max.
- Enable caching for long system prompts to get the $0.05 per million token rate.
- Set spending caps and a step limit for each computer-use session.
- Do not give the agent payment or deletion rights without a human approval step.
- Log every agent action (screenshots, clicks) so you can audit incidents.
Timeline and what has not been announced
- October 26, 2026: OpenAI announces GPT-6.1 Terra, opens the API and starts a gradual rollout in ChatGPT Work and Codex. OpenAI notes it may take a while to reach every account.
- No specific numbers yet on Terra usage limits in each ChatGPT plan.
- It is unclear whether, or when, OpenAI will retire GPT-5.6 Terra in the API. Teams on the old model should watch the deprecations page.
- No independent benchmark results yet; OSWorld 2.1 is the one to watch because it is closest to real computer use.
- Anthropic and Google have not responded on pricing; Claude Sonnet 5.5 currently matches GPT-6.1 Sol at $2/$10 (per InfoWorld).
To place Terra in a broader plan, see our guide to choosing and deploying AI agents for business, with cost tables. Further reading: GPT-6.1 Sol at $2/$10 compared with Astra and OpenAI's computer-use agent test on Ironclad contract software.
Sources checked: OpenAI – Introducing GPT-6.1 Terra, OpenAI API – gpt-6.1-terra model page, InfoWorld, DataCamp, Interesting Engineering.