GPT-6.1 Terra, launched by OpenAI on 26 October 2026, is a model built for office work handed to an AI: filling in forms, entering invoices, interpreting a drawing, operating applications on screen. It costs $1 per million input tokens and $5 output. GPT-6.1 Sol, OpenAI’s next model up, and Claude Sonnet 5.5, Anthropic’s mid-range model, are both listed at double that: $2 and $10.
On paper, the choice is easy. In practice, a model is paid by the task, not by the token, and the number of tokens it uses depends on how it reasons and on the effort level you ask of it. OpenAI publishes, for each of its tests, the score and cost per task of every model at every effort level, Claude included. That makes a precise comparison possible, and it does not produce the same winner at every setting.
The verdict in brief
- Listed price: Terra, half the price. $1 and $5 per million tokens, against $2 and $10 for Sol and for Sonnet 5.5 alike.
- Real cost per task: Terra, but not always by half. According to Artificial Analysis, $0.96 per task for Terra, $1.27 for Sol, $10.02 for Sonnet 5.5. On some of OpenAI’s tests, Terra costs as much as Sol for a lower score.
- Best results at max effort: Claude Sonnet 5.5. It has the best score on every OpenAI chart where it appears, and it is ahead of Terra and Sol at Artificial Analysis. But it costs up to ten times more per task.
- At medium effort, the default setting: Terra matches or beats Sonnet 5.5, for a cost per task between a third and nearly half lower, on OpenAI’s charts. Sol does better than both for a moderate extra cost.
- Complex coding: Sol or Sonnet 5.5. OpenAI itself advises against Terra for the most complex programming tasks.
- Speed: Terra. It completes its tasks fastest, at Artificial Analysis as at Plotly.
- Our advice: Terra for high volumes of office tasks and everyday screen control. Sol if you want a notch more reliability for a limited extra cost. Sonnet 5.5 when the best result matters more than the bill.
Where the figures in this comparison come from
We did not test these models ourselves. This comparison combines four types of sources, all cited:
- The charts in OpenAI’s announcement. The page publishes the exact values of six charts: score and average cost per task, for five effort levels, on computer-use, vision, coding and accuracy tests. Claude Sonnet 5.5 and Opus 5.5 appear on four of them. This is publisher data, measured by OpenAI.
- Independent measurements from Artificial Analysis, recorded on 30 October 2026: an index of nine tests, plus cost, time and tokens per task, all models at max effort.
- An independent exam from Plotly, a maker of data analytics tools: 43 private questions on a wind-turbine dataset.
- Feedback from the Hacker News launch thread (427 comments), quoted under each author’s public username.
One reading precaution, given by OpenAI itself: in that thread, tedsanders, who says they work at OpenAI, writes that the error bars on these charts are “commonly like ±5%”. A gap of two or three points between two models therefore does not name a winner.
The contenders, side by side
| GPT-6.1 Terra | GPT-6.1 Sol | Claude Sonnet 5.5 | Claude Opus 5.5 (reference) | |
|---|---|---|---|---|
| Publisher | OpenAI | OpenAI | Anthropic | Anthropic |
| API price (input / output, per million) | $1 / $5 | $2 / $10 | $2 / $10 | $4 / $20 |
| Artificial Analysis index | 51 | 55 | 54 | 58 |
| Average cost per task | $0.96 | $1.27 | $10.02 | $10.04 |
| Output tokens per task | 60,000 | 68,000 | 285,000 | 173,000 |
| Average time per task | 655 s | 861 s | 1,301 s | 1,152 s |
At OpenAI, GPT-6 Luna ($0.10 and $0.50) sits at the bottom of the range and GPT-6 Astra ($10 and $50) at the top; we use them as reference points where the charts show them. Terra accepts text and images as input, with a one-million-token context window and knowledge up to 1 June 2026.
The listed price, and the real price: cost per task
The table above already shows it: Sonnet 5.5 and Sol have the same per-token price, but Sonnet produces more than four times as many output tokens per task, and ends up eight times more expensive per task at Artificial Analysis. The listed price says almost nothing about the bill.
Between Terra and Sol, the listed gap is half. On OpenAI’s charts, at max effort, it breaks down as follows:
| Test (OpenAI's charts, max effort) | GPT-6.1 Terra | GPT-6.1 Sol | Cost-per-task gap |
|---|---|---|---|
| Agents' Last Exam (professional computer tasks) | 60.0% for $1.01 | 58.6% for $2.00 | −50% |
| Terminal-Bench-Science 0.1 (scientific research through code) | 58.3% for $4.15 | 57.0% for $5.47 | −24% |
| MMMU-Pro (visual reasoning) | 85.5% for $0.012 | 86.3% for $0.013 | −8% |
| Terminal-Bench 4.0 (command-line coding) | 50.3% for $1.84 | 55.9% for $1.98 | −7% |
| Factual errors (lower is better) | 4.3% for $0.12 | 4.4% for $0.12 | 0% |
| OSWorld 2.1 (desktop software) | 74.2% for $1.19 | 78.0% for $1.15 | +3% |
An almost identical cost per task with a per-token price half as high means that Terra consumed about twice as many tokens to complete the same task. On the Artificial Analysis index, Terra’s saving over Sol is 24%, not 50%.
Two conditions on Terra’s developer page also weigh on the bill: above 272,000 tokens sent in a request, the whole request is billed at twice the input rate and one and a half times the output rate; and regional processing, which keeps data in a given area, adds 10%.
The effort setting changes the winner
This is the data point that quick comparisons miss. Each model can be set to several reasoning effort levels, and the ranking is not the same at the bottom and the top of the scale. On the two OpenAI charts where Terra, Sol and Sonnet 5.5 appear together:
| OpenAI chart | GPT-6.1 Terra | GPT-6.1 Sol | Claude Sonnet 5.5 |
|---|---|---|---|
| OSWorld 2.1, medium effort | 66.7% for $0.50 | 73.5% for $0.65 | 65.9% for $0.93 |
| OSWorld 2.1, max effort | 74.2% for $1.19 | 78.0% for $1.15 | 83.9% for $5.74 |
| Terminal-Bench 4.0, medium effort | 35.8% for $0.46 | 52.7% for $1.02 | 28.8% for $0.68 |
| Terminal-Bench 4.0, max effort | 50.3% for $1.84 | 55.9% for $1.98 | 70.6% for $10.44 |
At max effort, Sonnet 5.5 clearly dominates, at a high price. At medium effort, the order flips: Terra is level with Sonnet on OSWorld and ahead of it on Terminal-Bench, for a cost per task between a third and nearly half lower. Medium effort is the default setting of Terra’s API, and the one Anthropic applies by default to Sonnet 5.5 in its apps and in Claude Code, as we noted in our Sonnet 5.5 vs Muse Wonder Full comparison. In other words, for a user who changes nothing, the comparison is far more favourable to Terra than max-effort rankings suggest.
In both cases, Sol remains OpenAI’s best compromise: first of the three at medium effort, for an extra cost of 30% to 122% over Terra depending on the test.
Computer use and office tasks: who wins
This is Terra’s claimed home ground. The results are mixed.
On Agents’ Last Exam, the test OpenAI puts forward, Terra is OpenAI’s best model: 60.0% at max effort, ahead of Sol (58.6%) and Astra (59.3%). At medium effort, it scores exactly the same as Astra (57.6%) for $0.29 per task against $4.82. The test includes tasks such as turning architectural drawings into a 3D model, or logging an animated film shot by shot for a restoration team. Claude Opus 5.5, the only Claude model shown on this chart, does better: 63.2%, for $8.82 per task.
On OSWorld 2.1, the reference test for using desktop applications, Terra is behind: 74.2% against 78.0% for Sol, 80.9% for Astra and 83.9% for Sonnet 5.5.
At Artificial Analysis, on the two tests closest to office work handed to an agent, Terra is the weakest of the four models compared: 60% on AutomationBench (tasks in online software), against 65% for Sol, 72% for Sonnet 5.5 and 70% for Opus 5.5; and 1,402 Elo points on AA-Briefcase, against 1,555, 1,814 and 1,801.
At Plotly, on 43 data analytics questions, Terra gets 40 right for $1.31 per full run, Sol gets a perfect score (43 out of 43) for $2.61, and Luna 38 for $0.13. Plotly describes it as the fastest model on its leaderboard of 35 among those scoring 40 or more, with a median of 15 seconds per question.
The reading of tedsanders, who works at OpenAI, sums up this picture well: Terra would be better than Astra and Sol at technical communication, “similar on vision and computer use (but much faster!)”, and a bit worse at most other things.
Coding: Sol or Sonnet 5.5, not Terra
OpenAI says so in its announcement: for the most complex coding tasks, Sol and Astra “remain the better choice”. The figures confirm it. On Terminal-Bench 4.0, at max effort, Terra solves 50.3% of tasks, Sol 55.9%, Astra 58.2% and Sonnet 5.5 70.6%, the latter for $10.44 per task.
Artificial Analysis narrows the gap with Sol: on its own run of Terminal-Bench 4.0, Terra and Sol are level, at 56% each, and Sonnet 5.5 scores 64%. On FrontierCode 2.0, the three are neck and neck (42%, 41% and 42%). For everyday code, Terra is enough; for a demanding project, Sol at medium effort offers the best value, and Sonnet 5.5 at high or max effort the best result. Our comparison Claude Opus 5.5 vs GPT-6 Sol covers the top of the range in detail.
Answer reliability and vision
Factual errors. On an internal OpenAI test made of questions designed to trigger errors, Terra at max effort gives about one answer in 23 containing at least one error (4.3%), level with Sol (4.4%) and Astra (4.2%). At medium effort, one answer in ten (10.1%) still contains at least one error, and one in six at low effort (16.7%). That is clear progress on GPT-5.6 Terra (53.6% at low effort), but Claude does not appear on this chart.
Knowledge. Independent knowledge tests are less favourable to Terra: 30 on the AA-Omniscience index, like Sonnet 5.5, against 42 for Sol and 48 for Opus 5.5; 47% on HLE-Diamond, against 53% for Sonnet 5.5 and 59% for Sol. For specialised information retrieval, Sol is the better choice.
Vision. On MMMU-Pro, Terra is almost at the level of the large models: 85.5% at max effort, against 86.3% for Sol and 86.7% for Astra. It reads neither video nor audio, however: it can operate editing software or analyse frames taken from a film, not open a video file.
Speed: Terra’s advantage
Terra completes a task in 655 seconds on average at Artificial Analysis, against 861 for Sol, 1,152 for Opus 5.5 and 1,301 for Sonnet 5.5. Sonnet actually writes faster (140 tokens per second against 91), but it writes much more. OpenAI announces Terra as “over three times faster” than Sol and Astra, on its own test; at Artificial Analysis, the gap with Sol is closer to 1.3 times. Plotly confirms the advantage: a median of 15 seconds per question for Terra, 18 for Sol.
What OpenAI’s charts do not say about Claude
The announcement cites Claude Opus 5.5 twice: to stress that Terra costs about nine times less per task on Agents’ Last Exam, and that it comes “within five points” of Opus on Terminal-Bench-Science. It does not say that on Agents’ Last Exam too, Opus has the best score. And it never cites Sonnet 5.5, even though it appears on two charts.
| Test (OpenAI chart) | Terra's best score | Best Claude score shown | Cost per task (Terra / Claude) |
|---|---|---|---|
| Agents' Last Exam | 60.0% | Opus 5.5: 63.2% | $1.01 / $8.82 |
| OSWorld 2.1 | 74.2% | Sonnet 5.5: 83.9% | $1.19 / $5.74 |
| Terminal-Bench 4.0 | 50.3% | Sonnet 5.5: 70.6% | $1.84 / $10.44 |
| Terminal-Bench-Science 0.1 | 58.3% | Opus 5.5: 63.3% | $4.15 / $23.21 |
On all four charts that include a Claude model, Claude has the best score against Terra; on three, it even beats every OpenAI model, GPT-6 Astra included. No Claude model appears on the factual-error chart or the vision chart. Readers noticed: on Hacker News, entrope asks why OpenAI does not show Opus 5.5 on OSWorld, nor any Anthropic model on the factual-error test. Finally, on Terminal-Bench-Science, Opus 5.5 appears “with fallbacks”: at Anthropic, some sensitive tasks are rerouted to an older model, which complicates the comparison.
Terra or Luna: is it worth paying ten times more?
GPT-6 Luna costs ten times less per token than Terra. On OpenAI’s charts, Terra at low effort roughly matches Luna at max effort, for a similar cost per task: 55.9% against 50.9% on Agents’ Last Exam ($0.16 against $0.14), 53.8% against 55.5% on OSWorld ($0.36 against $0.30). Terra pulls ahead as soon as it is given more effort: at max effort, 74.2% against 55.5% on OSWorld.
At Plotly, the gap is two correct answers out of 43 (40 against 38), for a run ten times more expensive ($1.31 against $0.13). For simple, highly repetitive tasks, Luna is often enough. One Hacker News user, frisbee6152, sums up the dilemma: when Luna is not enough, the jump to Sol is often worth it, which leaves little room for Terra. Plotly, by contrast, finds that Terra joins its cost-accuracy frontier: no model tested there is both cheaper and more accurate.
Agent safety: what Terra’s card says
For an agent that browses and acts on your behalf, resistance to prompt injection, instructions hidden in a page that try to divert it, is decisive. OpenAI’s system card puts figures on several behaviours. Anthropic does not publish the same tests for Sonnet 5.5, so a direct comparison with Claude is not possible here.
| Evaluation (OpenAI system card) | GPT-6.1 Terra | Comparison points |
|---|---|---|
| Resistance to indirect prompt injections | 98.77% | GPT-5.6 (June 2026): 96.23%; GPT-6.1 Sol: 98.98%; GPT-6 Astra: 99.79% |
| Getting around a legitimate restriction (lower is better) | 29.6% | GPT-5.6 Terra: 28.4%; GPT-6.1 Sol: 22.0%; GPT-6 Astra: 21.6% |
| Unauthorised action after messages from other agents | 13.75% (11 runs out of 80) | GPT-6.1 Sol: 1.25%; GPT-6 Sol: 16.25% |
Against injections, Terra does three times better than the previous generation, but remains below Sol and Astra: roughly one attack in 80 gets through. On the other hand, it gets around a legitimate restriction slightly more often than its predecessor, and OpenAI explains why: most often, the model mistakes a legitimate restriction for an injection attempt and ignores it. These tests cover low-stakes situations without the protections of the real system. On this front, Sol is the most cautious choice among OpenAI’s models. Like Sol, Terra is rated at the “Critical” level of OpenAI’s cybersecurity risk scale, which we covered in detail for GPT-6.1 Sol.
In ChatGPT: Work, not Chat, and a default that changes
According to OpenAI’s Help Center, ChatGPT is split into two spaces: Chat for quick questions, Work, an agent for longer, multi-step tasks that end in a deliverable, and alongside them Codex for development. GPT-6.1 Terra, like Sol, is offered only in Work and Codex, for the Plus, Pro, Business, Enterprise and Edu plans according to the release notes (the announcement does not list Business). No free plan is mentioned. OpenAI’s always-on agents, the dots, also launch their tasks in Work and Codex.
One detail in the release notes matters for this comparison: some Plus and Pro users will see GPT-6.1 Terra at medium effort instead of GPT-6.1 Sol at light effort in Work’s default presets. The model running your tasks may therefore have changed without you choosing it. The model picker lets you switch back to Sol.
What early users say
Four days after launch, public feedback is still limited, but it points to a consensus. The Hacker News launch thread, with 427 comments, sets the tone.
- The best use is medium effort. XCSme, who publishes their own tests, writes that at max effort Terra is “worse than 6.1 sol and 2x more expensive” in their runs, but that medium effort is its best use case, cheaper than Sol.
- A narrow place in the range. For minimaxir, Terra remains a “weird middle child” between Luna, which got much better, and Sol, which got much cheaper. Silagi goes further, for Terra and Sonnet alike: almost every task would be better served by going up a tier with less effort, or down a tier with more effort.
- Price per token is not enough. stymaar points out that the price ordering holds “per token, yes”, but not necessarily per task, “depending on how much thinking is needed”.
- A model on probation? tedsanders, who works at OpenAI, writes that the company is excited about Terra’s fast computer use, but that if its uptake is poor, OpenAI might go back to three models. For subscribers, the same message adds, it is “not a bombshell release”.
These are individual accounts, not measurements. They point in the same direction as the figures: Terra is a good tool for a specific use, not a universal replacement.
What our own ranking says
Below this article, an automatic summary drawn from our AI ranking compares the two models. It shows Claude Sonnet 5.5 winning every measured use, including value for money, which seems to contradict the costs per task quoted above. The explanation lies in the variants and the sources. The summary keeps, on each side, the best-ranked variant: for Sonnet 5.5, the Xhigh version, 9th in our ranking, but measured by a single source, LMArena, which aggregates blind user votes; for Terra, the Max version, 289th, measured by Artificial Analysis alone. The two figures therefore do not come from the same source, and the block flags this itself, for three of the five uses, with a caveat on the strength of the measurements. On a single source, Artificial Analysis, Sonnet 5.5 at max effort is still ahead of Terra, but by much less: 210th against 289th.
Which one to choose for your situation
| Your situation | Our pick | Why |
|---|---|---|
| High volumes of office tasks (forms, invoices, data entry) | GPT-6.1 Terra | OpenAI's best model on Agents' Last Exam, lowest cost per task after Luna |
| High-stakes control of desktop applications | GPT-6.1 Sol, or Sonnet 5.5 if the budget allows | On OSWorld, Sol beats Terra at the same cost, Sonnet beats them all |
| Complex coding | GPT-6.1 Sol or Claude Sonnet 5.5 | Terra is the weakest of the three on Terminal-Bench 4.0 according to OpenAI |
| Everyday use without adjusting effort | GPT-6.1 Terra | At medium effort, it matches or beats Sonnet for a clearly lower cost per task |
| Specialised information retrieval | GPT-6.1 Sol | Best knowledge scores of the three at Artificial Analysis |
| Simple, highly repetitive tasks | GPT-6 Luna | Ten times cheaper than Terra, almost at the same level at low effort |
| Autonomous agents that can pay or delete | GPT-6.1 Sol, with human approval | Better safety results than Terra in OpenAI's card |
For each camp’s most powerful models, our comparison GPT-6 Astra vs Fable 5.1 details the gaps in price and results.
The bottom line
GPT-6.1 Terra does not beat Claude Sonnet 5.5 on peak performance, and it does not replace Sol for demanding code. It does better than that for one specific use: office work at volume, at the default setting, where it matches Sonnet for a clearly lower cost per task and is faster than all of them. The lesson goes beyond this trio: a model should be judged on the cost of the finished task and on the effort level you will actually use, not on the price per million tokens at the top of the announcement.