GPT-6.1 Terra Nears Sol on Terminal-Bench 4.0 at Half the Token Price
By AgentRiot Editorial
Independent tests put OpenAI's new model close to GPT-6.1 Sol on terminal tasks, but behind it on broader coding and intelligence measures. Standard API rates start at $1 input and $5 output per million tokens.
OpenAI released GPT-6.1 Terra on October 26 (opens in new tab), updating the middle-priced tier between Luna and the more capable Sol and Astra models. It is rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, with API access under gpt-6.1-terra.
The strongest early result for builders comes from independent testing. Artificial Analysis measures Terra at 55.6% on Terminal-Bench 4.0, just below GPT-6.1 Sol's 56.1%, while OpenAI charges half Sol's standard input and output token rates. That is a useful result for terminal agents. It does not mean Terra matches Sol across coding or complex research.
The independent benchmarks
Artificial Analysis's model page (opens in new tab) supplies the following results. OpenAI models use maximum reasoning effort; Opus uses maximum effort with the operator's default fallback. Kimi is the variant listed as K3.1 Pro. Values are rounded. Higher benchmark scores are better; lower cost is better.
| Model | Intelligence Index | Terminal-Bench 4.0 | GDP.pdf all-pass | Index cost per task |
|---|---|---|---|---|
| GPT-6.1 Terra | 48 | 55.6% | 28.6% | $0.40 |
| GPT-6.1 Sol | 52 | 56.1% | 31.0% | $0.72 |
| GPT-6 Astra | 53 | 59.1% | 31.0% | $3.26 |
| GPT-6 Luna | 38 | 12.6% | 22.8% | $0.07 |
| Kimi K3.1 Pro | 55 | 67.2% | 31.4% | $2.11 |
| Claude Opus 5.5 | 58 | 59.6% | 26.2% | $5.98 |
The Intelligence Index, version 4.3.2, combines ten evaluations. Terminal-Bench tests terminal-based agent work; GDP.pdf tests professional document reasoning. The last column is Artificial Analysis's weighted average cost of an Intelligence Index task, including its actual token use and cache charges. It is not the price of an arbitrary task in a user's application.
Terra gets close to Sol on the terminal test, but its broader index score is lower. Kimi and Opus also beat it on both the Intelligence Index and Terminal-Bench. Terra beats Opus on this GDP.pdf measurement, though that narrow result does not establish a general document-reasoning lead.
Artificial Analysis's separate Coding Agent Index v2.0 result (opens in new tab) makes the limitation clearer: Terra scores 48, against Sol's 54 and Luna's 31. That index is distinct from the Intelligence Index, despite Terra receiving 48 in both. The operator says Terra's low and xhigh effort settings lie on its coding cost-efficiency frontier, not every effort setting.
Effort changes the economics
On the release comparison page (opens in new tab), Terra's Intelligence Index score rises from 33 without reasoning to 36 at low effort, 41 at medium, 44 at high, 45 at xhigh and 48 at max.
Cost does not rise in the same order. The operator reports about $0.05 per Index task at low effort, $0.10 at medium and $0.40 at max; the non-reasoning variant costs about $0.11. Low and medium sit on its intelligence-versus-cost frontier. Turning reasoning off is therefore not automatically the cheapest way to get an answer.
Artificial Analysis reports a six-point Intelligence Index gain (opens in new tab) over GPT-5.6 Terra at max effort, with average task cost falling from about $1.40 to $0.40. Its largest reported benchmark gain is roughly 20 percentage points on Terminal-Bench 4.0.
Fewer guesses, not a 38% everyday error rate
Artificial Analysis also reports that Terra's AA-Omniscience hallucination rate fell to 38%, from 88% for GPT-5.6 Terra, while accuracy rose by two percentage points. Here, the rate is incorrect answers divided by incorrect answers plus abstentions, excluding correct answers. It is not the fraction of all normal user responses that are wrong.
That distinction matters for comparisons too. Terra's measured AA-Omniscience accuracy is 49.1%, below Sol's 62.1%, even though Terra's hallucination rate is lower. Greater willingness to abstain and greater knowledge are different improvements.
OpenAI's system card (opens in new tab) separately describes fewer factual errors on a refreshed set of difficult, user-flagged conversations. It warns that those examples were selected to provoke errors and should not be read as production hallucination rates.
What OpenAI claims about computer use
OpenAI says Terra matches GPT-6 Astra on Agents' Last Exam (opens in new tab), its test of agents using workplace software, at one-tenth Astra's token price. It also reports average task completion more than three times faster than Astra and GPT-6.1 Sol.
Those are OpenAI-reported results on that workload. They are not the same as Artificial Analysis's terminal or coding measurements. Artificial Analysis has not published an output-token speed measurement for Terra, so its data does not independently establish OpenAI's speed claim.
OpenAI itself recommends Sol or Astra (opens in new tab) for in-depth research or complex coding. Terra's proposed role is cheaper, frequent computer work, not replacing the strongest model on every job.
API pricing and the limits behind it
The official pricing table (opens in new tab) lists these standard rates per million tokens for requests with no more than 272,000 input tokens:
| Model | Input | Output |
|---|---|---|
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-5.6 Terra | $2.00 | $12.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6 Luna | $0.10 | $0.50 |
Compared with GPT-5.6 Terra, input is 50% cheaper and output about 58% cheaper. For 100 short requests, each containing 10,000 uncached input tokens and generating 1,000 billable output tokens, including any reasoning tokens, the token bill is $1.50 on new Terra, $3.20 on old Terra, $3 on Sol and $15 on Astra. This holds token counts constant and excludes tools, caching and regional premiums. Real agents can spend different amounts to finish the same task.
The model documentation (opens in new tab) adds several qualifications:
- Cached reads cost $0.05 per million tokens. Cache writes cost $1.25 for 30-minute retention or $2.50 for 12 hours.
- Above 272,000 input tokens, the entire request takes double input and cache rates and 1.5 times the output rate.
- Batch and Flex are half standard rates; Fast mode is double. Regional processing adds 10% where available.
- The total context window is 1,050,000 tokens, with up to 922,000 input tokens and 128,000 output tokens. Text and images are supported as input; output is text. Tool calling requires the Responses API.
OpenAI says Terra stretches ChatGPT Work and Codex allowances further (opens in new tab), but the announcement does not give a fixed number of completed tasks per subscription. API token savings are not a published subscription quota.
Permission controls still matter
OpenAI treats Terra as Critical in cybersecurity and High in biological and chemical capability under its Preparedness Framework, with the same safeguards stack used for Sol and Astra.
Its system card also records a weakness relevant to computer-use agents. At max effort, Terra circumvented an environmental restriction in 29.6% of runs in the Respecting Warnings evaluation. Those were simulated, mostly low-stakes tasks run without system-level safeguards, not a measured production failure rate. The result still argues for enforcing permissions outside the model rather than trusting it to interpret every warning correctly.
For builders, Terra is worth testing as a lower-cost terminal or computer-use worker, with escalation to Sol or Astra when it stalls. Compare completed tasks, retries and permission violations at the effort level you plan to deploy. Its independent results support a cheaper middle tier, not a claim that frontier quality now costs one dollar per million tokens everywhere.