OpenAI released GPT-6.1 Terra on 2026-10-26, according to its API changelog, and priced it at half of GPT-6.1 Sol per token.[1][5] This Model Watch reads the API documentation for Terra, Sol, Astra and the older GPT-5.6 Terra side by side, works out what a sample workload costs on each, and sets out which one fits which job. We did not run any of these models, and everything below was read on 2026-11-01.
Quick answer
- What it is: OpenAI's changelog calls Terra "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing".[5]
- Price: $1 per million input tokens and $5 per million output tokens at standard rates, with cached input at $0.05.[1]
- How it compares: Sol is $2 and $10; Astra is $10 and $50; GPT-5.6 Terra is $2 and $12.[2][3][6]
- Who it is for, per OpenAI: visual reasoning, computer use and writing. For "the most demanding reasoning, coding, and professional work" the guide points to Astra.[4]
- What we could not check: quality. No documentation page tells you how much worse or better Terra is on your task. Only your own test does.
What the documentation says, side by side
| Spec | GPT-6.1 Terra | GPT-6.1 Sol | GPT-6 Astra | GPT-5.6 Terra |
|---|---|---|---|---|
| Input per 1M tokens | $1 | $2 | $10 | $2 |
| Cached input per 1M | $0.05 | $0.10 | $1 | $0.20 |
| Output per 1M tokens | $5 | $10 | $50 | $12 |
| Context window | 1,050,000 | 1,050,000 | 1,050,000 | 1,050,000 |
| Max output | 128,000 | 128,000 | 128,000 | 128,000 |
| Knowledge cutoff | 2026-06-01 | 2026-04-30 | 2026-04-30 | 2026-02-16 |
| Input / output | Text, image / text | Text, image / text | Text, image / text | Text, image / text |
All figures are the standard rate from the model pages.[1][2][3][6] Three things stand out.
The context window is identical. If you were choosing on how much you can send, there is no difference among the three GPT-6 models. The difference is price and capability, not capacity.
Terra has the newest knowledge cutoff. The Terra page lists 2026-06-01, against 2026-04-30 for Sol and Astra.[1][2][3] That is a small practical advantage for questions about recent events and library versions, and it is the one place the cheapest tier is ahead on paper.
The surcharge rules are documented on Sol and Astra, not on the Terra text we read. Sol's page says prompts above 272K input tokens cost 2x for input and cache and 1.5x for output across the full request, and the Astra page states the same rule.[2][3] The summary we got of the Terra page did not list a long-context multiplier. Do not assume there is none; read the live page before sending very long prompts.
What the price gap means for a real workload
The per-token ratios are easy to quote and easy to misapply. Here is a worked example, and it is our arithmetic, not an OpenAI figure.
Take 1,000 tasks, each sending 20,000 input tokens and receiving 2,000 output tokens, with no caching and no long-context surcharge. That is 20 million input tokens and 2 million output tokens.
| Model | Input cost | Output cost | Total |
|---|---|---|---|
| GPT-6.1 Terra | $20 | $10 | $30 |
| GPT-6.1 Sol | $40 | $20 | $60 |
| GPT-6 Astra | $200 | $100 | $300 |
| GPT-5.6 Terra | $40 | $24 | $64 |
On this mix, Terra costs half of Sol, a tenth of Astra and about 47% of GPT-5.6 Terra. Compared with GPT-5.6 Terra, the input price is 50% lower, the output price is about 58% lower ($5 against $12) and the cached input price is 75% lower ($0.05 against $0.20), all by our arithmetic from the model pages.[1][6]
Two caveats keep this from being a buying decision:
- Reasoning tokens. Reasoning models bill their thinking as output. A model that reasons longer on the same task can cost more than its price card suggests, and we have no documented figure for how many tokens each model uses on a given task. Our explainer on reasoning models, thinking tokens and cost covers how to measure this.
- Retries and review. If the cheaper model fails a task more often, the cost of a retry or a human fix can erase the saving. The price gap tells you the ceiling on the saving, not the saving.
Choosing between them
OpenAI's own guide gives the clearest signal. It describes Terra as "Fast and efficient. For visual reasoning, precise computer use, and clear, natural writing", Sol as "Near-Astra performance for complex work at a lower cost", and Astra as "For the most demanding reasoning, coding, and professional work".[4] It suggests tasks such as matching product photos to catalogue entries and acting on application screens for Terra.[4]
| If your task is... | Start with | Why (our reading of the documentation) |
|---|---|---|
| High-volume image or screen understanding | Terra | OpenAI positions it for visual reasoning and computer use, at the lowest price[4][1] |
| Drafting, rewriting, summarising at scale | Terra | Positioned for clear, natural writing; cost matters most at volume[4] |
| Complex coding or multi-step professional work | Sol, then test Astra | Sol is described as near-Astra at lower cost; Astra is the vendor's top tier[4][2] |
| The hardest reasoning where errors are expensive | Astra | The vendor's most capable model; ten times Terra's price per token[3][4] |
| A short, simple call where latency and cost dominate | Terra with a low or no reasoning effort | Terra accepts the none reasoning effort[4] |
The method matters more than the table: run twenty to fifty of your own real tasks through Terra first, score them, and move up a tier only for the task types that fail. This is our recommendation, not OpenAI's. For model selection beyond OpenAI, see our GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini pricing comparison.
Migration notes that change your code
The guide documents three practical points.[4]
- Reasoning effort. Terra supports none, low, medium (the default), high, xhigh and max. Sol and Astra do not support none. OpenAI suggests starting at medium and raising effort only when the quality gain justifies the extra time and tokens.[4]
- Moving from GPT-5.6 Terra. Compare both models on representative images and workflows at the same reasoning effort, so you are not measuring a settings change.[4]
- Tools. If you call tools through Chat Completions, the guide says to migrate those requests to the Responses API.[4] The changelog also lists Astra's migration limits: no custom temperature or top_p, no logprobs, and tool calling through the Responses API.[5] We did not verify whether those same limits apply to Terra; check before relying on sampling parameters. For how the mechanics work across vendors, see how tool calling works in Claude, OpenAI and Gemini.
The changelog also shows how fast this tier moves: GPT-5.6 Terra launched on 2026-07-09 and had its price reduced by 20% on 2026-07-30.[5] Treat any price in this article as a dated reading and re-check before budgeting.
What remains uncertain
- Quality. We found no OpenAI documentation that quantifies Terra against Sol or Astra on a task type, and we did not test. Secondary coverage of benchmark and error-rate claims exists, but we did not open OpenAI's announcement or system card (the pages returned an access error to our fetcher), so we do not repeat those numbers.
- Safety classification. Reports say OpenAI rates the Terra family highly under its preparedness framework for cybersecurity. We have not read the system card ourselves. If you are giving Terra tools or computer access, read the card's agent-safety and cyber sections first; our guide to reading an AI system card lists what to look for.
- Long-context pricing on Terra and any rate limits for your account tier.
Frequently asked questions
What is GPT-6.1 Terra? OpenAI's changelog describes it as a fast, efficient model for visual reasoning, precise computer use and clear, natural writing, released 2026-10-26.[5]
How much does it cost? $1 per million input tokens, $0.05 cached input and $5 per million output tokens at standard rates; Batch and Flex are 50% lower.[1]
Is it cheaper than GPT-6.1 Sol? Yes, half the per-token price: Sol lists $2 and $10.[1][2]
Does it replace Astra? Not according to OpenAI's guide, which still points the most demanding work to Astra.[4]
Can I turn reasoning off? The guide says Terra supports none as a reasoning effort; Sol and Astra do not.[4]
