OpenAI has released GPT-6.1 Terra, a mid-tier model aimed at fast, large-scale computer use, and cut its standard API price to $1 per million input tokens and $5 per million output tokens. The Decoder and Techstrong.ai both report the new rates, which sit well below the previous Terra. For teams running agents that read screens, click and retry, the GPT-6.1 Terra price matters because agent loops resend large contexts again and again.
OpenAI
AnthropicWhat GPT-6.1 Terra is and where it is available
The API model name is gpt-6.1-terra, per the OpenAI developer community announcement. Techstrong.ai describes it as an update to OpenAI's mid-tier model, placed between the larger Sol and Astra models and the lowest-cost Luna tier, and designed for computer-based work at lower cost.
Availability, according to the same sources: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, plus the Responses API, which supports computer use, function calling, file search and external tool connections. OpenAI's own announcement page returned an access error when we tried to open it, so details here rely on the secondary coverage and the community post.
GPT-6.1 Terra pricing: input halves, output falls further
| Model | Input | Output |
|---|---|---|
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-5.6 Terra (after July cut) | $2.00 | $12.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
USD per million tokens, standard rates. Sources: The Decoder and Techstrong.ai (Terra versions), Anthropic pricing (Sonnet 5.5), OpenAI model page (Sol).
Headlines say "half the price," but the cut is uneven. Input falls 50% ($2 to $1), while output falls from $12 to $5, a reduction of about 58% (Techstrong.ai). The Decoder adds that cached input is $0.05 per million tokens, which it describes as a quarter of the previous rate. Business Today and OpenAI's community post report the same three figures, and OpenAI says per-token cost is one-tenth of GPT-6 Astra's.
Standard API list price per million tokens
Input and output prices in USD for the Terra versions and nearby models. Bars share one linear scale, so Astra dwarfs the rest.
Benchmarks: what OpenAI and Artificial Analysis report
OpenAI's reported numbers, as relayed by Techstrong.ai and the community post:
- Agents' Last Exam (agents using workplace software), medium effort: 15 percentage points above GPT-5.6 Terra at roughly one-third the cost per task, and finishing more than three times faster than Astra and Sol.
- Terminal-Bench 4.0 (command-line coding agents), maximum effort: 50.3%, against 21.5% for GPT-5.6 Terra. OpenAI says Sol and Astra remain better for the most complex coding.
Independent figures from Artificial Analysis, via The Decoder: an Intelligence Index of 48, against 42 for GPT-5.6 Terra and 52 for GPT-6.1 Sol. At maximum effort the Index gains six points while average cost per task falls from $1.40 to $0.40.
Input is exactly half, but output falls from $12 to $5, a larger cut.
The Decoder, on GPT-6.1 Terra pricing
Hallucination: the evidence is narrower than the headlines
Social posts have circulated specific hallucination-rate drops for Terra. We could not confirm them in any of the sources above, so we do not repeat them. What The Decoder does report, from the AA-Omniscience test, is that the model guesses 38% of the time when it lacks an answer, down from 88% for GPT-5.6 Terra, with accuracy up two points. That measures guessing behavior, not an overall error rate in everyday use. Business Today also advises reviewing important agent actions rather than assuming tasks complete correctly.
Agent-loop economics: what the price gap means
Agents pay for tokens repeatedly. Each step re-sends accumulated screenshots, tool output and instructions, so input volume usually dwarfs output. The table below is an illustration, not a measurement: it assumes one hypothetical task that consumes 1.0 million input tokens and 0.1 million output tokens, using list prices only. Real tasks vary by harness, effort setting and how many tokens a model spends.
| Model | No caching | 80% of input cached |
|---|---|---|
| GPT-6.1 Terra | $1.50 | $0.74 |
| GPT-5.6 Terra | $3.20 | n/a |
| Claude Sonnet 5.5 | $3.00 | $1.56 |
| GPT-6.1 Sol | $3.00 | $1.48 |
Arithmetic on the list prices above, with cached reads at $0.05 (Terra), $0.20 (Sonnet 5.5, per Anthropic) and $0.10 (Sol, per OpenAI). Cache-write charges are ignored, and GPT-5.6 Terra's cached rate was not confirmed.
On these assumptions GPT-6.1 Terra costs about half as much per task as the others when nothing is cached, and the gap widens once most context is re-read from cache. Token efficiency can move the picture either way: Artificial Analysis's measured per-task figure ($0.40 against $1.40 for the previous Terra) reflects both price and token use, which a list-price table cannot capture. Comparisons across vendors also hinge on quality at the task level, where Sonnet 5.5 and Sol may lead on some benchmarks (see our Sonnet 5.5 and GPT-6.1 Sol coverage).
What this means for model choice
- Price per token is no longer the main variable. With a mid-tier model at $1 and $5, the choice shifts to task success rate and tokens used per task.
- Pricing moves quickly. Terra's own list price changed twice in about three months, down to $2 and $12 in July and now to $1 and $5, according to Techstrong.ai.
- Portability reduces lock-in. Teams that can switch models per task can follow these changes. Metir is a model-agnostic workspace where chat and agent work can run on models from several providers, so a price change at one vendor does not require rebuilding a workflow.
Sources:
- GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger models - The Decoder
- OpenAI's GPT-6.1 Terra improves computer use, cuts API prices - Techstrong.ai
- GPT-6.1 Terra: delivering fast and capable computer use at a lower price - OpenAI Developer Community
- OpenAI launches GPT-6.1 Terra at half the API price - Business Today
- Claude pricing - Anthropic
- GPT-6.1 Sol model page - OpenAI API docs
Image credits
Header image: the OpenAI website seen through a magnifying glass, a 2023 stock-style photo by Jernej Furman, Wikimedia Commons, CC BY 2.0 (source); it is not an image of the GPT-6.1 Terra announcement. In-body image: screenshot of the OpenAI Codex desktop app by Wikideas1, Wikimedia Commons, CC0 (source); it shows a GPT-5.6-generation model, not GPT-6.1 Terra.