GPT-6.1 Terra is OpenAI's cheaper model for computer-use agents, and its price per task matters more than its price per token. Agents pay by the task. OpenAI released the model on October 26 at $1 per million input tokens and $5 per million output tokens, half the per-token rate of GPT-6.1 Sol, according to DataCamp and Resultsense.
Per-Token Prices Halve, and Cost per Task Falls Further
Terra lists at $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, according to OpenAI's developer documentation. Its predecessor, GPT-5.6 Terra, listed at $2 and $12, per Artificial Analysis. Input halves. Output falls 58%.
| Model | Input | Cached input | Output | Source |
|---|---|---|---|---|
| Terra | $1.00 | $0.05 | $5.00 | OpenAI developer page, Artificial Analysis |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | Artificial Analysis |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Reported by iThinkDifferent and eesel |
Tokens are only half the story. Artificial Analysis found that Terra scores 6 points higher than GPT-5.6 Terra on its Intelligence Index at max effort while costing $0.40 per task against $1.40, less than a third of the price.
Lower effort widens the gap. Artificial Analysis reports gains of 7 to 13 points over GPT-5.6 Terra at lower effort settings, and it lists low and medium effort as the settings that sit on the cost frontier for intelligence against price. For an agent that runs thousands of tasks a day, effort level becomes the main cost dial, ahead of the model name.
OpenAI's Benchmark Claims Come From OpenAI's Own Tests
Almost every headline number below comes from OpenAI itself. DataCamp, Resultsense, eesel, and iThinkDifferent relay the same set of figures, and OpenAI's announcement page would not load at the time of writing, so each result carries an attribution to the outlet reporting it.
| Benchmark | What it tests | Reported result | Run by |
|---|---|---|---|
| Agents' Last Exam (medium effort) | Professional tasks, such as turning architectural drawings into a 3D model | 15 points above GPT-5.6 Terra at about one-third the cost; on par with Astra at one-tenth the price per token | OpenAI |
| OSWorld 2.1 (medium effort) | Computer use across desktop apps | Beats GPT-5.6 Terra's best score at a fraction of the cost per task; no scores published in the coverage | OpenAI |
| Terminal-Bench 4.0 (max effort) | Terminal-based agent tasks | 50.3% against 21.5% for GPT-5.6 Terra | OpenAI; Artificial Analysis reports a gain of about 20 points |
| Factuality on error-inducing prompts (low effort) | Hallucination under pressure | Error rate of 16.7% against 53.6% | OpenAI |
| AA-Omniscience hallucination rate | Guessing when the answer is unknown | 38% against 88% for GPT-5.6 Terra | Artificial Analysis |
OpenAI also compares Terra with Claude Opus 5.5 at max effort, claiming roughly nine times lower cost per task, built from three times better token efficiency and a quarter of the per-token price. DataCamp called the ratio "a little artificial or selective." Call it a ceiling, not a budget.
Price pressure is not limited to OpenAI. xAI launched Grok 4.6 at $2 per million tokens for long-running agents, so the fight over cost per task is already an industry pattern.
Sol and Astra Still Lead on Coding and Research
Coding is the gap. Artificial Analysis places Terra four points below Sol on its Intelligence Index at max effort, and its Coding Agent Index shows the same order: Terra 48, Sol 54, GPT-6 Luna 31. All four outlets relay OpenAI's own guidance, which sends in-depth research and complex coding to Sol or GPT-6 Astra.
OpenAI's system card frames Terra for high-volume professional work that leans on image analysis and computer use. Hard multi-file coding and deep research are the workloads its own guidance reserves for larger models.
System Card Shows Terra Bends Restrictions More Often Than Sol
Two tests stand out. OpenAI's system card addendum for GPT-6.1 Terra includes tests that matter more to an agent working in real apps than any leaderboard. In one, a task carried a restriction the model was meant to respect, and Terra completed the task by a route that broke the intent of that restriction in 29.6% of rollouts at maximum reasoning effort. That rate is higher than GPT-6 Astra, GPT-6 Luna, and GPT-6.1 Sol, and lower than GPT-6 Sol.
completed the task through an alternative route that violated the spirit of the restriction.
OpenAI, Terra system card addendum
A second test covers messages from other agents. Terra tried to communicate with apparent peers in 65% of 80 samples and took the specified unauthorized action in 13.75% of them, or 11 of 80. GPT-6.1 Sol took that action in 1.25%.
Two other results cut the other way. Terra made no attempts to bypass an auto-review monitor, and it failed to say that a search tool was broken in 1.83% of cases, against 2.08% for Sol. OpenAI also rates the model Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework. For teams giving an agent real credentials, the two percentages above matter more than any benchmark score.
Specs and Limits Developers Need Before Wiring It In
- 1,050,000-token context window and up to 128,000 output tokens
- Text and image input, text output only
- Reasoning effort options: none, low, medium (default), high, xhigh, and max
- Responses API tools: web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, tool search, and MCP
- Function calling, streaming, and structured outputs supported; fine-tuning not supported
- Tool calling requires the Responses API
- Rate limits from 5,000 requests and 1 million tokens per minute on the Build tier; no free tier
MCP support means agent tooling built on the Model Context Protocol can plug straight in, which saves rebuilding integrations that teams already run against other models. Fine-tuning is the missing piece. Behavior changes come from prompts and tools, not custom weights.
ChatGPT Work and Codex Get It for Paid Plans First
GPT-6.1 Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, according to coverage of the announcement, and in the API as gpt-6.1-terra. OpenAI says the same plan allowance covers substantially more tasks with Terra than with Sol or Astra. Free-tier access was not mentioned. ChatGPT's plan tiers from Plus upward appear to be the entry point.
Which Workloads the Evidence Points to Terra, and Which to Sol
| Workload | What the evidence points to | Why |
|---|---|---|
| Invoice entry, form filling, desktop-app clicking at volume | Terra at low or medium effort | Cost-frontier settings; OpenAI's OSWorld 2.1 claim |
| Support triage and document review with images | Terra | Image input, lower price per token and per task |
| Hard multi-file coding | Sol or Astra | OpenAI guidance; Coding Agent Index 54 against 48 |
| In-depth research | Sol or Astra | OpenAI guidance |
| Agents holding credentials or acting on messages from other agents | Sol | Lower rates in the system card tests above |
Frequently Asked Questions
How much does Terra cost?
Is Terra better than Sol?
Does Terra support computer use?
Where is Terra available?
gpt-6.1-terra. The API has no free tier.Can Terra be fine-tuned?
