GPT-6.1 Terra Review: $1/$5 Built for Computer Use
Quick verdict
GPT-6.1 Terra is OpenAI's efficiency-tier model for visual reasoning, computer use and writing, released October 26, 2026 at $1 input / $5 output per million tokens — exactly half of GPT-6.1 Sol on every line. It pairs a 1,050,000-token context (922K input, 128K output) with reasoning effort from none to max, but tool calling requires the Responses API, and requests above 272K input tokens double input prices for the entire request. OpenAI says it matches GPT-6 Astra at roughly a tenth of the token price; at maximum effort it scores 48 on the Artificial Analysis Intelligence Index for about $0.40 per task, so Sol at medium still wins on cost per point.
What Shipped
OpenAI released GPT-6.1 Terra on October 26, 2026, as the efficiency tier of the GPT-6.1 generation. There was never a GPT-6 Terra — the line jumps from GPT-5.6 Terra to 6.1 — and the price falls a long way on the jump: $1 input and $5 output per million tokens, against $2/$12 for GPT-5.6 Terra and $2/$10 for GPT-6.1 Sol.
The positioning is narrow on purpose. Terra is built for visual reasoning, precise computer use and natural writing — matching product photos to catalog entries, reading technical drawings, operating application screens, filling forms, drafting prose. It is not pitched as a frontier coding model; OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better choice for the hardest coding work.
It ships in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, not in the regular ChatGPT chat, and through the API as gpt-6.1-terra. Amazon Bedrock added it the same day.
Specs at a Glance
| Spec | GPT-6.1 Terra |
|---|---|
| Model ID | gpt-6.1-terra |
| Context window | 1,050,000 tokens |
| Maximum input | 922,000 tokens |
| Maximum output | 128,000 tokens |
| Modalities | Text and image in; text out |
| Knowledge cutoff | June 1, 2026 |
| Reasoning effort | none, low, medium (default), high, xhigh, max |
| Tool calling | Responses API required |
| Fine-tuning | Not supported |
| Availability | ChatGPT Work, Codex, API, Amazon Bedrock |
Two details matter more than the headline context number. First, reasoning.effort accepts none — a setting GPT-6.1 Sol dropped — so Terra can run as a fast non-reasoning model for extraction and routing. Second, tool calling is Responses-only, which changes integration work for anyone currently on Chat Completions.
Pricing: Half of Sol, With a 272K Cliff
| Per 1M tokens (Standard) | GPT-6.1 Terra | GPT-6.1 Sol | GPT-5.6 Terra |
|---|---|---|---|
| Input | $1.00 | $2.00 | $2.00 |
| Cached input | $0.05 | $0.10 | $0.20 |
| Output | $5.00 | $10.00 | $12.00 |
| Cache write, 30-minute | $1.25 | $2.50 | $2.50 |
| Fast mode | $2.00 / $10.00 | $2.00 / $10.00 | — |
The rate card is exactly half of Sol, and Terra's Fast tier lands precisely on Sol's standard card. Batch and Flex are another 50% down.
Then there is the long-prompt adjustment: above 272,000 input tokens, input and cached-input rates double and output rises 1.5x, applied to the entire request rather than the tokens above the line. A request that crosses the threshold at 280K tokens bills as if every input token cost $2. The lesson is operational, not theoretical: keep retrieval and summarization workloads under the line, or accept that a 1.05M-token window is a capability you rarely want to pay for.
Benchmarks: OpenAI's Numbers, and What They Mean
OpenAI's published comparisons are the only ones available so far:
- Terminal-Bench 4.0: 50.3% at maximum effort, against GPT-5.6 Terra's published 21.5%.
- Agents' Last Exam: 15 percentage points above GPT-5.6 Terra at medium effort, at roughly one-third of the cost, according to OpenAI.
- Against GPT-6 Astra: OpenAI says Terra performs on par at about one-tenth the per-token price, and finishes tasks more than three times faster.
- Internal error-rate test: error-inducing prompts at low effort fell from 53.6% with GPT-5.6 Terra to 16.7%; at maximum effort the rate was 4.3%. The test set is internal and cannot be inspected.
Independent numbers are thinner but useful. Artificial Analysis scores Terra at 48 on its Intelligence Index at maximum effort, with a Coding Agent Index v2.0 score of 48 against Sol's 54 and Luna's 31. Its cost-per-task table is the more interesting read: Terra at high effort scores 44 for about $0.13 per task, two points above Sol at low effort for the same money, while Terra at max costs $0.40 for the same 48 points Sol reaches at medium for $0.21.
That is the shape of the model: excellent value in the 40–45 point band, ordinary value at the top of its effort range.
Where Terra Fits
The GPT-6.1 lineup now has three tiers with clear jobs:
- Luna ($0.10/$0.50): routing, classification, high-volume extraction.
- Terra ($1/$5): visual reasoning, computer use, structured document work, writing where quality matters but frontier reasoning does not.
- Sol ($2/$10) and Astra ($10/$50): hard coding, long-horizon agentic work, anything where the last few index points decide the outcome.
The practical rule from the cost data: run Terra at high effort for most agent steps, drop to low or none for extraction, and escalate to Sol medium when a task actually needs 48+ points — it is cheaper per point than Terra at max.
Before You Switch
Four checks decide whether Terra is a drop-in replacement or a migration project:
- Tool calling. Chat Completions still works for plain text, but any tool loop must move to the Responses API. Teams with production function-calling code should budget for that rewrite.
- Effort defaults. Medium is the default; low and none are available for extraction and routing. Since cost scales with generated tokens, the effort setting matters more than the model choice for simple tasks.
- The 272K threshold. Keep retrieval, summarization and repository-context workloads under it, or the whole request bills at double input rates.
- Cache retention. Cache writes cost $1.25 for 30 minutes or $2.50 for 12 hours. Agent sessions that pause between turns usually justify the longer retention; one-shot batch jobs do not.
Limitations
- Vendor-run benchmarks. Terminal-Bench and Agents' Last Exam figures come from OpenAI; no independent reproduction of the 50.3% result exists yet.
- The 272K cliff. Long-context requests are priced punitively, which undercuts the practical value of the 1.05M window.
- Responses-only tool calling. Existing Chat Completions tool loops need rework.
- No audio or video input. Terra is text-and-image only.
- Not a frontier coder. OpenAI's own guidance points complex coding to Sol or Astra.
Who Should Use GPT-6.1 Terra
- Computer-use agent builders who need screen reading, form filling and bounded interface workflows at a fraction of Astra's token price.
- Document and catalog teams doing visual matching, technical drawing comparison and structured extraction at volume.
- Writing-heavy products that want better prose than Luna without Sol's bill.
- Not for the hardest coding tasks, audio or video pipelines, or workloads that routinely exceed 272K input tokens.
Final Verdict
GPT-6.1 Terra is the most useful model OpenAI has shipped for automation teams this quarter, and the least glamorous: no leaderboard crown, no frontier claims, just half of Sol's rate card attached to a genuinely useful multimodal skill set.
Two numbers summarize it. At high effort, 44 index points for about $0.13 per task is the best value OpenAI sells. At max effort, 48 points for $0.40 is worse than Sol at medium. Use the effort dial deliberately and Terra earns its place; leave it on max and you are paying frontier prices for a mid-tier score.
Related Articles
Frequently Asked Questions
How much does GPT-6.1 Terra cost?+
How does GPT-6.1 Terra compare with GPT-6.1 Sol?+
What is GPT-6.1 Terra's context window?+
Is GPT-6.1 Terra available in ChatGPT?+
What is GPT-6.1 Terra best at?+
Does GPT-6.1 Terra support tool calling?+
Pros
- Half of GPT-6.1 Sol's rate card on every line: $1 input / $5 output / $0.05 cached input
- 1,050,000-token context with a 128,000-token output cap and a June 1, 2026 knowledge cutoff
- Reasoning can be switched off entirely — GPT-6.1 Sol dropped the 'none' setting
- Terminal-Bench 4.0 at 50.3% versus GPT-5.6 Terra's 21.5% (OpenAI-reported)
- Full Responses API tool surface for screen-reading and bounded computer-use workflows
Cons
- At max effort it ties Sol medium's index score (48) at nearly twice the cost per task
- Prompts above 272K input tokens double input-side rates and raise output 1.5x for the whole request
- Tool calling requires migrating from Chat Completions to the Responses API
- Text and image input only — no audio or video
- Headline benchmark claims are OpenAI-run; independent reproductions are not yet available
Keep reading