🔍 Read the full analysis: GPT-6.1 Terra May Help Teams Cut Costs On Routine AI Tasks on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI released GPT-6.1 Terra on October 26, 2026, positioning it as a lower-priced model for routine workplace and computer-based tasks. Artificial Analysis found its low- and medium-effort settings on the cost-efficiency frontier in its tests, but Terra scores below several more capable models and its highest setting costs more for each additional performance gain.
OpenAI released GPT-6.1 Terra on October 26, a model it describes as fast and efficient for routine workplace and computer-based tasks, with API prices set at $1 per million input tokens and $5 per million output tokens. Independent benchmarker Artificial Analysis found that Terra’s low- and medium-effort settings sit on its cost-efficiency frontier, but the model does not match several higher-scoring alternatives on its Intelligence Index.
Terra is available in the API as gpt-6.1-terra and to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, according to the launch report. The listed API rates are half those of GPT-5.6 Terra and lower than GPT-6.1 Sol’s $2 per million input and $10 per million output tokens. Cached input is listed at $0.05 per million tokens. OpenAI said the same ChatGPT allowance can cover substantially more tasks with Terra than with Sol or Astra.
Artificial Analysis gave Terra a 48.1 Intelligence Index score at maximum effort, with an average cost of $0.40 per task across its evaluation set. In the same comparison, GPT-6.1 Sol scored 51.8 at $0.72 per task, while Claude Opus 5.5 scored 57.6 at $5.98. These are benchmark results, not guarantees of what a particular company will pay: actual costs depend on task, model setting and usage.
The report says Artificial Analysis measured six effort settings, ranging from non-reasoning to maximum effort. Across those settings, scores ran from 33.5 to 48.1 and average task costs from about $0.05 to $0.40. Low and medium effort were on the benchmark’s cost-efficiency frontier; settings above them cost more without improving the score enough to remain on that frontier. Terra’s non-reasoning mode was not the cheapest per task: the reported average was $0.107, compared with $0.055 at low effort.
Infographic 01 · The effort curve
One model, six settings, a 7-fold cost range.
GPT-6.1 Terra at each effort level, as measured by Artificial Analysis. Low and medium sit on its cost-efficiency frontier: no tested model is both cheaper and better.
Score · Intelligence Index
independent higher is better
Cost per finished task
note the last step shorter is cheaper
All six settings
| Effort | Index | Cost / task | Output tokens / task | Hallucination rate |
|---|---|---|---|---|
| Non-reasoning | Index 33.5 | Cost $0.107 | Tokens 4,737 | Hallucination 72% |
| Low | Index 36.2 | Cost $0.055 | Tokens 2,681 | Hallucination 46% |
| Medium | Index 41.3 | Cost $0.095 | Tokens 5,835 | Hallucination 44% |
| High | Index 43.9 | Cost $0.134 | Tokens 9,027 | Hallucination 43% |
| Xhigh | Index 45.4 | Cost $0.194 | Tokens 14,551 | Hallucination 42% |
| Max | Index 48.1 | Cost $0.403 | Tokens 35,096 | Hallucination 38% |
Hallucination rate in AA-Omniscience: how often the model answers anyway when it does not know. Lower is better.
The first steps up are cheap. The last one doubles the bill.Assessment: Thorsten Meyer
Sources: Artificial Analysis model pages (retrieved 26 Oct 2026); Artificial Analysis launch article (26 Oct 2026). Cost per task is the average across the index’s evaluation tasks.
Powered by Thorsten Meyer AI
Why Task Costs Change the Calculation
The launch highlights a difference between token price and the cost of finishing a task. A model with a lower rate per token may produce more tokens, while a more expensive model may complete a job with fewer. Artificial Analysis found that GPT-6 Luna’s listed token rates were lower than Terra’s, yet Terra at low effort had a lower average task cost in the index. That result applies to the evaluated tasks, not necessarily every workload.
For businesses handling high volumes of routine work, such as invoice processing, the benchmark offers a reason to test a smaller or cheaper model before routing every request to a more capable one. OpenAI itself gives invoice reading and entry into accounting software as an example of Terra’s intended use. Companies still need to measure accuracy, review needs and total costs on their own data; a low benchmark cost does not establish that a workflow can run without human oversight.
The performance trade-off is material. Terra’s maximum-effort score is below Sol’s and far below Opus 5.5’s in the cited comparison. The report also notes that an open-weight model, Xiaomi’s MiMo-V2.6-Pro, scored 46.3 for an average $0.13 per task, placing it ahead of Terra’s higher effort settings on both reported score and cost. The choice is therefore not simply between Terra and OpenAI’s largest models.
Products Worth Considering
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Terra’s Place in OpenAI’s Lineup
OpenAI describes Terra as a model for communication and computer-based work, rather than as its top-performing option. The source report places it between the smaller GPT-6 Luna and larger Sol and Astra models. OpenAI’s launch information says Terra is more than three times faster than Sol and Astra, but Artificial Analysis had not published speed measurements for Terra at the time covered by the report. That speed comparison remains an OpenAI claim.
OpenAI also cited results from its own tests. On Agents’ Last Exam, it reported Terra at medium effort scoring 57.6%, compared with 42.6% for GPT-5.6 Terra, at a stated task cost of $0.29 versus $0.84. OpenAI said Terra matched GPT-6 Astra’s medium-effort score on that test, with a lower stated cost. On Terminal-Bench 4.0, OpenAI reported a 50.3% score for Terra at maximum effort, versus 21.5% for GPT-5.6 Terra. These are company-reported comparisons and should be read separately from Artificial Analysis’s independent index.
OpenAI’s launch material also reports lower error rates on an internal set of prompts designed to induce factual errors. It gave rates of 16.7% at low effort and 4.3% at maximum effort for Terra, compared with 53.6% and 22.3% for GPT-5.6 Terra at the corresponding settings. The source does not provide enough detail here to judge how well that internal test represents errors in everyday deployments.
Limits of the Benchmark Evidence
The results do not establish how Terra will perform or cost in any one company’s workflow. Artificial Analysis’s task costs are averages across its index evaluations, and the source does not provide enough information to translate them into a guaranteed price for invoice processing, coding or other specific work. Performance can also vary by prompt, workload and selected effort level.
Some comparisons come from OpenAI’s own tests, while others come from Artificial Analysis; they use different evaluations and should not be treated as interchangeable. The available material also does not include independent speed measurements, despite OpenAI’s claim that Terra is more than three times faster than Sol and Astra. It remains unclear how the reported speed, error and cost results will hold up across broader use.
The report says Terra guessed on 38% of questions it could not answer in Artificial Analysis’s hallucination test, down from 88% for GPT-5.6 Terra. That is a result on a particular benchmark, not evidence that Terra will avoid unsupported answers in general. OpenAI’s statement that Terra meets the “Critical” threshold for cybersecurity under its Preparedness Framework is also the company’s own classification.
Testing Terra in Real Workflows
The next meaningful test is whether the benchmark advantage carries over to real production workloads. Teams considering Terra can compare it with their current model on representative tasks, recording not only API charges but also completion rates, correction time, latency and the cost of human review. Those results will show whether low or medium effort is sufficient, or whether a higher setting or a stronger model is needed.
Further independent speed measurements and evaluations across practical workplace tasks would help clarify OpenAI’s launch claims. Until those results are available, Terra’s cost profile is best treated as a reason to run controlled trials—not as proof that it will reduce costs for every organization. OpenAI’s own guidance, as reported, is that Sol and Astra remain preferable for the hardest coding tasks.
Key Questions
What is GPT-6.1 Terra?
GPT-6.1 Terra is a new OpenAI model marketed for routine communication and computer-based work. It is available through the API and, according to the launch report, in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.
How much does GPT-6.1 Terra cost?
The reported API prices are $1 per million input tokens and $5 per million output tokens, with cached input at $0.05 per million tokens. Artificial Analysis measured an average cost of $0.40 per evaluation task at maximum effort and $0.055 at low effort; those averages are not fixed prices for every task.
Is Terra cheaper than GPT-6 Luna?
Not by token rate: the report lists Luna at lower per-token prices. But in Artificial Analysis’s evaluated tasks, Terra at low effort had a lower average cost per task than Luna at maximum effort. The comparison depends on the benchmark workload and should not be assumed to apply to every use.
Does Terra match OpenAI’s larger models?
No such conclusion is supported by the reported results. Terra scored below GPT-6.1 Sol and Claude Opus 5.5 on Artificial Analysis’s cited maximum-effort comparison. OpenAI also says Sol and Astra remain better choices for its most complex coding tasks.
What remains unproven about Terra’s cost savings?
The benchmark does not establish savings for a particular business. Teams still need to test Terra on their own tasks and include accuracy, corrections, latency and human review in the cost calculation. Independent speed measurements were not available in the report.
Source: ThorstenMeyerAI.com
