# GPT-6.1 Terra: close to Opus in one test, nearly 9× lower cost

> Official benchmarks compare Terra with Sol, Astra and Claude. Explore reasoning effort, task costs, API pricing and access in ChatGPT Work and Codex.


Published: 2026-10-26. Original source: [OpenAI · Introducing GPT-6.1 Terra](https://openai.com/index/introducing-gpt-6-1-terra/).

![GPT-6.1 Terra · OpenAI](/news/gpt61-terra/official-cover.webp)

Official launch image from [OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/). Interactive globe by HeyMark using official textures optimized for the web.

**In a professional computer-use test, GPT-6.1 Terra scores 60% at US$1.01 per case. Claude Opus 5.5 scores 63.2% at US$8.82.** Those are OpenAI’s **Agents’ Last Exam v1 results, both at Max**: Terra trails by 3.2 percentage points at roughly 8.7 times lower published cost. That trade-off is the striking part of the [GPT-6.1 Terra launch](https://openai.com/index/introducing-gpt-6-1-terra/).

If you are choosing a model for work, the question is practical: how much performance do you need, and what does it cost to get it? Terra’s position changes with the task. Explore the six tests to see where the trade-off works for you.

OpenAI positions Terra as a model for **understanding what is on screen and acting inside apps**, alongside writing. Tasks in the launch evaluation include turning drawings into a 3D model and reviewing a film shot by shot. These are the provider’s examples, not demos run here. Standard API pricing is **50% lower than GPT-5.6 Terra and GPT-6.1 Sol**.

## GPT-6.1 Terra benchmarks, with their costs

The chart contains **all 143 original data points** from the launch charts. Each symbol represents a reported configuration. Switch tests, compare by cost or reasoning effort, and pick a level to read the numbers. The complete data is underneath, including models with a result at Max only.

## GPT-6.1 Terra benchmarks: complete data
143 original configurations from OpenAI’s charts, October 26, 2026. Published USD per evaluated case; distinct from prices per million tokens. Lines connect reported configurations; effort settings are model-specific, without interpolating missing values.

### Agents’ Last Exam v1

Professional computer-use tasks

Score: higher is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| Claude Opus 5.5 | Max | **63.2 %** | US$8.82 |
| GPT-5.6 Terra | Low | 40.2 % | US$0.59 |
| GPT-5.6 Terra | Medium | 42.6 % | US$0.84 |
| GPT-5.6 Terra | High | 46.6 % | US$1.72 |
| GPT-5.6 Terra | Xhigh | 48.5 % | US$2.50 |
| GPT-5.6 Terra | Max | 50.7 % | US$3.58 |
| GPT-6 Astra | Low | 53.4 % | US$3.55 |
| GPT-6 Astra | Medium | 57.6 % | US$4.82 |
| GPT-6 Astra | High | 57.8 % | US$5.15 |
| GPT-6 Astra | Xhigh | 58.3 % | US$5.91 |
| GPT-6 Astra | Max | 59.3 % | US$7.24 |
| GPT-6 Luna | Low | 36.3 % | US$0.022 |
| GPT-6 Luna | Medium | 46.8 % | US$0.093 |
| GPT-6 Luna | High | 43.6 % | US$0.087 |
| GPT-6 Luna | Xhigh | 47.9 % | US$0.093 |
| GPT-6 Luna | Max | 50.9 % | US$0.14 |
| GPT-6.1 Sol | Low | 52.4 % | US$0.70 |
| GPT-6.1 Sol | Medium | 53.8 % | US$1.31 |
| GPT-6.1 Sol | High | 55.3 % | US$1.45 |
| GPT-6.1 Sol | Xhigh | 56.5 % | US$1.45 |
| GPT-6.1 Sol | Max | 58.6 % | US$2.00 |
| GPT-6.1 Terra | Low | 55.9 % | US$0.16 |
| GPT-6.1 Terra | Medium | 57.6 % | US$0.29 |
| GPT-6.1 Terra | High | 57.2 % | US$0.39 |
| GPT-6.1 Terra | Xhigh | 58.5 % | US$0.66 |
| GPT-6.1 Terra | Max | 60.0 % | US$1.01 |

### MMMU-Pro (Vision)

Visual reasoning

Score: higher is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| GPT-5.6 Terra | Low | 74.9 % | US$0.011 |
| GPT-5.6 Terra | Medium | 76.4 % | US$0.012 |
| GPT-5.6 Terra | High | 77.5 % | US$0.014 |
| GPT-5.6 Terra | Xhigh | 78.7 % | US$0.02 |
| GPT-5.6 Terra | Max | 81.7 % | US$0.079 |
| GPT-6 Astra | Low | 85.3 % | US$0.038 |
| GPT-6 Astra | Medium | 85.3 % | US$0.039 |
| GPT-6 Astra | High | 86.4 % | US$0.044 |
| GPT-6 Astra | Xhigh | 86.5 % | US$0.051 |
| GPT-6 Astra | Max | **86.7 %** | US$0.065 |
| GPT-6 Luna | Low | 73.5 % | US$0.0005 |
| GPT-6 Luna | Medium | 75.1 % | US$0.0006 |
| GPT-6 Luna | High | 76.5 % | US$0.0007 |
| GPT-6 Luna | Xhigh | 77.9 % | US$0.001 |
| GPT-6 Luna | Max | 79.9 % | US$0.0021 |
| GPT-6.1 Sol | Low | 84.0 % | US$0.0076 |
| GPT-6.1 Sol | Medium | 84.7 % | US$0.0079 |
| GPT-6.1 Sol | High | 85.8 % | US$0.0088 |
| GPT-6.1 Sol | Xhigh | 85.8 % | US$0.01 |
| GPT-6.1 Sol | Max | 86.3 % | US$0.013 |
| GPT-6.1 Terra | Low | 83.7 % | US$0.0038 |
| GPT-6.1 Terra | Medium | 84.6 % | US$0.0041 |
| GPT-6.1 Terra | High | 85.0 % | US$0.0045 |
| GPT-6.1 Terra | Xhigh | 85.4 % | US$0.0055 |
| GPT-6.1 Terra | Max | 85.5 % | US$0.012 |

### OSWorld 2.1 · offline

Desktop applications

Score: higher is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| Claude Sonnet 5.5 | Low | 57.8 % | US$0.68 |
| Claude Sonnet 5.5 | Medium | 65.9 % | US$0.93 |
| Claude Sonnet 5.5 | High | 73.3 % | US$1.39 |
| Claude Sonnet 5.5 | Xhigh | 81.0 % | US$2.22 |
| Claude Sonnet 5.5 | Max | **83.9 %** | US$5.74 |
| GPT-5.6 Terra | Low | 24.1 % | US$0.44 |
| GPT-5.6 Terra | Medium | 27.4 % | US$0.59 |
| GPT-5.6 Terra | High | 37.7 % | US$1.19 |
| GPT-5.6 Terra | Xhigh | 48.4 % | US$2.25 |
| GPT-5.6 Terra | Max | 57.1 % | US$4.76 |
| GPT-6 Astra | Low | 63.3 % | US$2.32 |
| GPT-6 Astra | Medium | 75.4 % | US$4.37 |
| GPT-6 Astra | High | 78.3 % | US$5.76 |
| GPT-6 Astra | Xhigh | 79.0 % | US$6.60 |
| GPT-6 Astra | Max | 80.9 % | US$8.17 |
| GPT-6 Luna | Low | 7.5 % | US$0.02 |
| GPT-6 Luna | Medium | 26.3 % | US$0.079 |
| GPT-6 Luna | High | 39.7 % | US$0.14 |
| GPT-6 Luna | Xhigh | 44.6 % | US$0.19 |
| GPT-6 Luna | Max | 55.5 % | US$0.30 |
| GPT-6.1 Sol | Low | 60.9 % | US$0.38 |
| GPT-6.1 Sol | Medium | 73.5 % | US$0.65 |
| GPT-6.1 Sol | High | 75.5 % | US$0.90 |
| GPT-6.1 Sol | Xhigh | 75.9 % | US$0.95 |
| GPT-6.1 Sol | Max | 78.0 % | US$1.15 |
| GPT-6.1 Terra | Low | 53.8 % | US$0.36 |
| GPT-6.1 Terra | Medium | 66.7 % | US$0.50 |
| GPT-6.1 Terra | High | 68.7 % | US$0.51 |
| GPT-6.1 Terra | Xhigh | 71.3 % | US$0.62 |
| GPT-6.1 Terra | Max | 74.2 % | US$1.19 |

### Factuality · difficult prompts

Answers with any factual error

Factual error: lower is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| GPT-5.6 Terra | Low | 53.6 % | US$0.055 |
| GPT-5.6 Terra | Medium | 51.5 % | US$0.061 |
| GPT-5.6 Terra | High | 41.1 % | US$0.078 |
| GPT-5.6 Terra | Xhigh | 33.9 % | US$0.11 |
| GPT-5.6 Terra | Max | 22.3 % | US$0.25 |
| GPT-6 Astra | Low | 13.3 % | US$0.26 |
| GPT-6 Astra | Medium | 9.7 % | US$0.29 |
| GPT-6 Astra | High | 5.9 % | US$0.43 |
| GPT-6 Astra | Xhigh | 4.6 % | US$0.55 |
| GPT-6 Astra | Max | **4.2 %** | US$0.66 |
| GPT-6 Luna | Low | 44.3 % | US$0.0025 |
| GPT-6 Luna | Medium | 31.2 % | US$0.0035 |
| GPT-6 Luna | High | 27.6 % | US$0.0047 |
| GPT-6 Luna | Xhigh | 24.5 % | US$0.0061 |
| GPT-6 Luna | Max | 18.4 % | US$0.013 |
| GPT-6.1 Sol | Low | 14.7 % | US$0.048 |
| GPT-6.1 Sol | Medium | 12.0 % | US$0.056 |
| GPT-6.1 Sol | High | 6.6 % | US$0.079 |
| GPT-6.1 Sol | Xhigh | 5.5 % | US$0.099 |
| GPT-6.1 Sol | Max | 4.4 % | US$0.12 |
| GPT-6.1 Terra | Low | 16.7 % | US$0.024 |
| GPT-6.1 Terra | Medium | 10.1 % | US$0.029 |
| GPT-6.1 Terra | High | 7.7 % | US$0.036 |
| GPT-6.1 Terra | Xhigh | 5.0 % | US$0.05 |
| GPT-6.1 Terra | Max | 4.3 % | US$0.12 |

### Terminal-Bench 4.0

Coding tasks in the terminal

Score: higher is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| Claude Sonnet 5.5 | Low | 20.0 % | US$0.62 |
| Claude Sonnet 5.5 | Medium | 28.8 % | US$0.68 |
| Claude Sonnet 5.5 | High | 43.0 % | US$1.46 |
| Claude Sonnet 5.5 | Xhigh | 61.5 % | US$4.34 |
| Claude Sonnet 5.5 | Max | **70.6 %** | US$10.44 |
| GPT-5.6 Terra | Max | 21.5 % | US$5.25 |
| GPT-6 Astra | Low | 50.6 % | US$4.93 |
| GPT-6 Astra | Medium | 54.4 % | US$6.06 |
| GPT-6 Astra | High | 58.1 % | US$7.18 |
| GPT-6 Astra | Xhigh | 57.9 % | US$7.45 |
| GPT-6 Astra | Max | 58.2 % | US$10.34 |
| GPT-6.1 Sol | Low | 47.1 % | US$0.76 |
| GPT-6.1 Sol | Medium | 52.7 % | US$1.02 |
| GPT-6.1 Sol | High | 53.0 % | US$1.28 |
| GPT-6.1 Sol | Xhigh | 53.4 % | US$1.28 |
| GPT-6.1 Sol | Max | 55.9 % | US$1.98 |
| GPT-6.1 Terra | Low | 24.8 % | US$0.24 |
| GPT-6.1 Terra | Medium | 35.8 % | US$0.46 |
| GPT-6.1 Terra | High | 41.8 % | US$0.65 |
| GPT-6.1 Terra | Xhigh | 46.4 % | US$0.99 |
| GPT-6.1 Terra | Max | 50.3 % | US$1.84 |

### Terminal-Bench-Science 0.1

Scientific research using code

Score: higher is better.

| Model | Effort | % | USD |
| --- | --- | ---: | ---: |
| Claude Opus 5.5 w/ fallbacks | Max | 63.3 % | US$23.21 |
| GPT-6 Astra | Low | 55.4 % | US$11.41 |
| GPT-6 Astra | Medium | 57.4 % | US$12.34 |
| GPT-6 Astra | High | 62.0 % | US$14.95 |
| GPT-6 Astra | Xhigh | 60.9 % | US$15.76 |
| GPT-6 Astra | Max | **68.1 %** | US$23.80 |
| GPT-6.1 Sol | Low | 43.7 % | US$1.79 |
| GPT-6.1 Sol | Medium | 47.6 % | US$2.34 |
| GPT-6.1 Sol | High | 51.1 % | US$2.76 |
| GPT-6.1 Sol | Xhigh | 53.7 % | US$2.89 |
| GPT-6.1 Sol | Max | 57.0 % | US$5.47 |
| GPT-6.1 Terra | Low | 24.3 % | US$0.47 |
| GPT-6.1 Terra | Medium | 34.3 % | US$0.95 |
| GPT-6.1 Terra | High | 39.7 % | US$1.24 |
| GPT-6.1 Terra | Xhigh | 43.4 % | US$1.72 |
| GPT-6.1 Terra | Max | 58.3 % | US$4.15 |

OSWorld uses the 2.1 offline set. Opus in Science uses fallbacks. Work/Codex allowances and latency cannot be inferred from these costs. This is not independent HeyMark testing or an average across benchmarks.
[OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)

At **Medium** on Agents’ Last Exam, Terra and Astra both score **57.6%**. Their published costs are **US$0.29 and US$4.82**, respectively. Matching this result does not make the models interchangeable for every assignment.

More reasoning does not always improve a result, either: Terra goes from **57.6% at Medium to 57.2% at High** on this test. The chart preserves that dip. A point’s cost is not a fixed tariff you can apply to any task of your own.

## Where other models still lead

Switching tabs changes the picture. All these figures are at **Max**, within the named test:

- **Vision, MMMU-Pro:** Terra scores **85.5%**, Sol **86.3%** and Astra **86.7%**. Terra is close, but it does not lead this comparison.
- **Desktop work, OSWorld 2.1 offline:** Terra reaches **74.2%**, versus Sol’s **78%** and Sonnet 5.5’s **83.9%**. The dataset name matters when comparing these figures with other OSWorld rankings.
- **Coding, Terminal-Bench 4.0:** Terra scores **50.3%**, versus **21.5%** for Terra 5.6. Sol reaches **55.9%** and Sonnet 5.5 **70.6%**. If difficult coding is your priority, inspect this tab before choosing from the first chart.
- **Scientific work with code:** Terra scores **58.3% at US$4.15**, ahead of Sol’s **57% at US$5.47**. The Opus result uses **fallbacks**: **63.3% at US$23.21**. That is a five-point gap, and the additional configuration is part of the comparison.

**HeyMark has not run these tests.** These are results published by OpenAI; every point and full model label is available in the table. The launch also reports task completion more than three times faster than Astra and Sol, but individual timings cannot be inferred from the USD values on the charts.

## Where is Opus 5.5? Beyond the launch curves

**Opus appears in two of [OpenAI’s six charts](https://openai.com/index/introducing-gpt-6-1-terra/).** The Claude competitor in OSWorld and Terminal-Bench 4.0 is Sonnet 5.5; vision and factual errors include no Claude model. In the chart specifications, OpenAI assigns Claude grey in the original, alongside green for Terra and blue for other GPTs. We give it its own color and a visible label in both themes.

Does the selection help sell the story of a cheap, competitive Terra? **That is an editorial hypothesis, not demonstrated intent.** Available runs or configuration differences could also affect the selection. We found no public explanation in the announcement for each omission.

External Opus results do exist for vision, desktop use and coding. Explore them below. We did not find an equivalent public Opus run for the internal factual-error test.

### Opus: what is in the curves and what is missing
Opus appears in 2 of the 6 original charts. The other four contain no Opus result; two include Sonnet. Checked October 26, 2026.
| Benchmark | Opus 5.5 | Other Claude |
| --- | --- | --- |
| Agents’ Last Exam v1 | Included | — |
| MMMU-Pro | Absent | — |
| OSWorld 2.1 · offline | Absent | Sonnet 5.5 |
| Factual errors | Absent | — |
| Terminal-Bench 4.0 | Absent | Sonnet 5.5 |
| Terminal-Bench-Science 0.1 | Included | — |

### MMMU-Pro

Artificial Analysis reports 87.69% (87.7% rounded) on MMMU-Pro for Opus 5.5 Max, Default Fallback. We did not find Terra 6.1 in its list at our cutoff. This is not a matched run against the 85.5% in OpenAI’s curve.
[Artificial Analysis](https://artificialanalysis.ai/evaluations/mmmu-pro)

### OSWorld

Anthropic reports 81.8% partial for Opus 5.5 Max; its system card also gives 48.7% strict. These are different metrics. The card describes 108 tasks and five runs, labeling its section OSWorld 2.0 while the launch says 2.1. Its setup is not equivalent to OpenAI’s offline set.
[Anthropic](https://www.anthropic.com/claude-opus-5-5) · [System card](https://cdn.sanity.io/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdf)

### Terminal-Bench 4.0

Four published Opus results, four configurations. We do not average them: agent, effort, repetitions and fallbacks can change the result. We did not find Terra 6.1 evaluated alongside Opus under the same protocol in these sources.

| Source | Opus 5.5 | Configuration |
| --- | --- | --- |
| [Artificial Analysis](https://artificialanalysis.ai/evaluations/terminalbench-4-0) | 59.6% | Max · Default Fallback · mini-swe-agent |
| [Terminal-Bench](https://www.tbench.ai/?efforts=all&version=4.0) | 64.8% ±3.1 pp | Max · Claude Code |
| [Vals AI](https://www.vals.ai/benchmarks/terminal-bench-4) | 65.15% | Mini-SWE-agent · avg@3 · 24 Oct |
| [Anthropic](https://www.anthropic.com/claude-opus-5-5) | 66.4% ±2.6 pp | Xhigh · production safeguards / fallbacks |

Comparison within Vals’ leaderboard, updated October 24: Mini-SWE-agent, 66 tasks, three runs. Terra 6.1 is not listed yet. The table does not detail effort or fallback policy for the current Opus result.

| Model | % | USD / test |
| --- | ---: | ---: |
| Claude Opus 5.5 | **65.15%** | $13.20 |
| GPT-6 Astra | 59.60% | $9.58 |
| GPT-6.1 Sol | 55.05% | $1.72 |

Vals’ earlier model page retains a different result: 61.62%, or 53.54% counting fallbacks as failures (30/198 assisted cases). We do not apply that adjustment to the current 65.15%: the source does not explain the difference.
[Vals: current leaderboard](https://www.vals.ai/benchmarks/terminal-bench-4) · [Vals: earlier model page](https://www.vals.ai/models/anthropic_claude-opus-5-5)

Missing from a chart does not mean missing a result. But bringing in a number from another website does not make a fair head-to-head. **We found no independent run testing the new Terra against Opus with the same task set and configuration.** That is why sources stay separate: Opus’ 87.7% vision result is not pasted onto Terra’s curve as though they sat the same exam.

## What the factual-error reduction measures

The internal difficult-prompt evaluation measures **answers containing any factual error**. At Low, Terra 5.6 records **53.6%** and Terra 6.1 **16.7%**. At Max, the new model falls to **4.3%**, versus Astra’s **4.2%** and Sol’s **4.4%**.

These are error rates: **lower is better**. They do not mean Terra gets 4.3% of everything anyone asks it wrong. When preparing content, it still makes sense to check names, dates and claims against their sources.

## API pricing: US$1 input and US$5 output

The [official API model page](https://developers.openai.com/api/docs/models/gpt-6.1-terra) identifies it as **`gpt-6.1-terra`**. Standard rates are in USD per **one million tokens**:

| Token type | Price |
| --- | --- |
| Input | **US$1.00** |
| Cached input reads | **US$0.05** |
| Cache writes, 30 minutes | US$1.25 |
| Cache writes, 12 hours | US$2.50 |
| Output | **US$5.00** |

For illustration, **10,000 input tokens and 2,000 output tokens cost US$0.02** at standard input/output rates: `(10,000 × 1 + 2,000 × 5) / 1,000,000`. This is not a forecast of a task’s token use and excludes tools, fast modes and additional cache charges.

Long context has an important condition: **above 272,000 input tokens, the entire request is charged at 2× input/cache rates and 1.5× output rates**. The docs list a **1,050,000-token API context window** and up to **128,000 output tokens**. Those figures do not establish the limits inside every Work or Codex experience.

## Where it is available at launch

OpenAI announces Terra in **ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu**, as well as the API. That does not confirm availability in the regular ChatGPT model picker. The announcement says the same usage allowance covers more tasks with Terra, without publishing a numerical quota we can turn into messages or assignments.

**Original ChatGPT announcement:** [open the launch post on X](https://x.com/ChatGPT/status/2114781101822714241). The numbers and conditions in this story come from OpenAI’s full launch announcement and API documentation.

For a particular workflow, look at the test closest to your work: operating apps, interpreting images and programming measure different things. The [system-card addendum](https://deploymentsafety.openai.com/gpt-6-1-terra) also explains that research/API evaluations can differ from production because of tools, instructions and effort. Cost gives you a reason to consider Terra; the differences between tasks help you decide when to choose it.

## Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.

[Start free](https://app.heymark.ai) · [See how it works](https://heymark.ai/en/#product)

## Developer and agent resources

- [HeyMark MCP documentation](https://heymark.ai/en/mcp/)
- [llms.txt](https://heymark.ai/llms.txt)
- [Full site content for language models](https://heymark.ai/llms-full.txt)
- MCP protocol endpoint: `POST https://mcp.heymark.ai`
- OAuth protected-resource metadata: [/.well-known/oauth-protected-resource](https://mcp.heymark.ai/.well-known/oauth-protected-resource)
- MCP server card: [/server-card](https://mcp.heymark.ai/server-card)