GPT-6.1 Terra: price, performance and fit
Updated October 27, 2026 · first published October 27, 2026
OpenAI released GPT-6.1 Terra on October 26, 2026. Its standard API rate is $1 per million input tokens and $5 per million output tokens. That is half the input price and less than half the output price of GPT-5.6 Terra. The useful FinOps question is whether your workflow can complete with fewer model dollars after caching, reasoning, computer-use steps, retries, and human review are counted. Terra is positioned for visual reasoning and computer use; it is not automatically the cheapest fit for every request.
One naming detail matters: the new model is gpt-6.1-terra, called GPT-6.1 Terra. It is separate from GPT-6.1 Sol, GPT-6 Luna, and GPT-6 Astra. OpenAI’s release announcement and API model page describe the intended use and current specifications.
GPT-6 releases: price and workload fit
The sequence moved from the premium flagship toward lower-cost and more specialized work. The table uses standard API text-token prices per million tokens. It is a rate-card comparison, not a ranking of cost per successful task.
| Release | Model | Input / output | Best starting point |
|---|---|---|---|
| Sep 3 | GPT-6 Astra | $10 / $50 | Highest-stakes or hardest work where evaluation shows the quality is needed |
| Sep 22 | GPT-6 Sol | $2 / $10 | Complex professional work and coding at a lower rate than Astra |
| Sep 22 | GPT-6 Luna | $0.10 / $0.50 | Focused, high-volume tasks that pass quality checks |
| Sep 29 | GPT-6.1 Sol | $2 / $10 | Updated Sol for complex work; cached input is $0.10 per million |
| Oct 26 | GPT-6.1 Terra | $1 / $5 | Visual reasoning, screen-based computer use, and scoped work |
| Previous generation | GPT-5.6 Terra | $2 / $12 | Baseline for existing Terra traffic and migration tests |
GPT-6.1 Terra’s standard input rate is half the GPT-5.6 Terra and GPT-6.1 Sol rate; its output rate is half GPT-6.1 Sol’s and 58% below GPT-5.6 Terra’s. GPT-6 Luna remains much cheaper for simple text jobs. Astra remains the expensive flagship. That makes Terra a potential middle path for tasks where screen understanding or its measured capability matters. See OpenAI’s GPT-6 Sol and Luna announcement and GPT-6.1 Sol announcement for the release context. The API pricing page is the source of truth for rates and pricing modes.
What a migration could save
Consider one million tasks, each using 1,000 uncached input tokens and 500 output tokens. At standard rates, token charges would be:
| Model | Modeled spend | Compared with GPT-5.6 Terra |
|---|---|---|
| GPT-5.6 Terra ($2 / $12) | $8,000 | Baseline |
| GPT-6.1 Terra ($1 / $5) | $3,500 | 56.25% lower |
| GPT-6 Luna ($0.10 / $0.50) | $350 | 95.6% lower |
| GPT-6.1 Sol ($2 / $10) | $7,000 | 12.5% lower |
| GPT-6 Astra ($10 / $50) | $35,000 | 4.4x the baseline |
This is arithmetic from published list rates, not a forecast: it assumes the same token counts, uncached input, no retries, no tools, and no changes in task success. Luna’s low token bill does not make it the right choice if it produces more failures or escalations. Compare cost per accepted outcome and quality, not just price per token.
The details that change the invoice
| Price component | GPT-6.1 Terra rate or rule | FinOps implication |
|---|---|---|
| Cached input | $0.05 per million tokens | 5% of uncached input; measure cache hit rate by workload |
| Cache writes | $1.25 per million tokens at 30-minute retention; $2.50 at 12-hour retention | Include write charges when estimating whether long-lived caching pays off |
| Batch and Flex | 50% below Standard | Candidate for work that can tolerate asynchronous or flexible processing |
| Fast mode | 2x Standard | Use only when the latency benefit is worth the premium |
| Long context | Above 272K input tokens, the full request uses 2x input/cache and 1.5x output rates | There is a threshold jump, not a gradual surcharge; alert before crossing it |
| Regional processing and FedRAMP endpoints | 10% uplift where applicable | Model the endpoint premium in regulated-workload forecasts |
| Tools | Some tool use has separate charges | Attribute tool fees separately from model tokens |
At the long-context threshold, Terra’s effective rates become $2 input, $0.10 cached input, and $7.50 output per million tokens for the full request. A request just over 272K can therefore cost materially more than one just under it, even if the extra context is small. The API model page lists a 1.05-million-token context window and 128,000-token output limit; a large context window is capacity, not a reason to send the whole history on every turn.
What changed beyond the price
OpenAI positions Terra around visual reasoning, computer use, and clearer writing. It accepts image input, but not audio or video. Its model page lists reasoning effort from none through max, with medium as the default, a June 1, 2026 knowledge cutoff, and support for computer use and other Responses API tools.
OpenAI reports that, at medium effort, Terra scored 15 percentage points above GPT-5.6 Terra on Agents’ Last Exam at about one-third the cost per task. The announcement also says Terra at maximum effort was over three times as token-efficient as Claude Opus 5.5, with about nine times lower cost per task, and completed tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol. On OpenAI’s internal difficult-prompt factual-error test, its low-effort error rate was 16.7%, compared with 53.6% for GPT-5.6 Terra. Its maximum-effort Terminal-Bench 4.0 score was 50.3%, versus 21.5% for GPT-5.6 Terra.
These are OpenAI-reported evaluations, not independent guarantees or your production results. Benchmark task mixes, effort settings, tools, and scoring rules matter. The launch itself says Sol and Astra remain better choices for the most complex coding tasks. Treat the numbers as a reason to run a migration evaluation, not as a committed savings percentage.
A FinOps migration test that answers the real question
- Choose one workflow. Start with a bounded visual or computer-use process, such as extracting invoice fields for review. Do not start with an autonomous payment or vendor-bank-change flow.
- Replay the same cases. Use representative screens and documents, including missing fields, layout changes, access errors, and timeouts. Hold tools, prompts, and acceptance rules constant.
- Record the full unit cost. Log model ID, effort, uncached and cached tokens, cache writes, tool fees, screenshots or other metered inputs, retries, and human-review time.
- Compare cost per accepted task. Include failed attempts and review. Also compare completion rate, error severity, latency, and exception handling.
- Roll out with a guardrail. Set a quality floor, a per-task spend limit, an exception route, and a rollback path. Recheck actual spend after traffic ramps.
The decision rule is simple: route a task to Terra if it meets the quality and safety bar and reduces total cost per accepted result, or unlocks a useful screen-based workflow at an acceptable cost. Route simpler work to Luna when it passes the same bar. Keep Sol or Astra for cases where your own evaluation proves their extra capability is needed.
Sources: OpenAI: GPT-6.1 Terra; Terra API model page; API pricing; GPT-6 Sol and Luna; GPT-6.1 Sol. Prices and product details checked October 27, 2026.
Related
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →