TL;DR
- GPT-6.1 Terra (
gpt-6.1-terra) on HiAPI costs $0.25 per million input tokens and $1.25 per million output tokens, a quarter of OpenAI's list price of $1 / $5. - One order-extraction call (about 250 input and 200 output tokens) costs about $0.0003, so $1 covers roughly 3,200 calls.
- It's live on HiAPI: open the model page to try it in the browser. New accounts get up to 2,000 credits ($1).
Price table
| Token type | Total input ≤ 272K | Total input > 272K | OpenAI list (≤ 272K) |
|---|---|---|---|
| Input | $0.25 | $0.50 | $1.00 |
| Cached read | $0.0125 | $0.025 | $0.05 |
| Cache write | $0.3125 | $0.625 | — |
| Output | $1.25 | $1.875 | $5.00 |
HiAPI prices per million tokens, checked October 29, 2026. The model page shows credits; 2,000 credits = $1.
Reasoning tokens are counted inside output tokens and cached tokens inside input tokens, so you pay for each token once, based on the usage the API reports. A request that ends in response.failed is not billed.
The 272K line switches the whole request
When a single request's total input goes past 272,000 tokens, every token in that request, including input, output and cache, moves to the long-context rate. It isn't only the excess that costs more.
| One request | Input rate | Output rate | Cost of this request |
|---|---|---|---|
| 272,000 in + 2,000 out | $0.25 / M | $1.25 / M | $0.0705 |
| 272,001 in + 2,000 out | $0.50 / M | $1.875 / M | $0.14 |
One extra token nearly doubles the bill. If you send whole documents, strip tables of contents, headers, footers and duplicate appendices to stay under 272K.
How reasoning effort changes the bill
Terra has six reasoning efforts: none, low, medium (default), high, xhigh and max. We sent the same requirements-breakdown prompt (split a bulk price update feature into tasks, acceptance criteria and risks) once at each effort:
| Effort | Output tokens | Reasoning | Time | Cost |
|---|---|---|---|---|
| none | 2,871 | 0 | 113 s | $0.0036 |
| low | 2,834 | 50 | 66 s | $0.0036 |
| medium | 3,055 | 79 | 126 s | $0.0039 |
| high | 1,561 | 132 | 26 s | $0.0020 |
| xhigh | 2,012 | 181 | 35 s | $0.0025 |
| max | 9,281 | 7,250 | 156 s | $0.0116 |
gpt-6.1-terra, English prompt of 126 input tokens, one run per effort, October 29, 2026.
From none to xhigh, cost tracked the length of the answer, not the effort level. At max, reasoning jumped to 7,250 tokens and the run cost three to six times as much as the others. On the Chinese version of the same prompt, xhigh also went past 8,000 reasoning tokens.
Keep the default medium for everyday work. Use none or low for bulk formatting, classification and extraction. Save xhigh and max for hard problems. To control cost, put a length limit in the prompt ("each bullet under 25 words"); it does more than switching efforts.
What caching saves
Send a prompt_cache_key in the body and the same value in a session_id header, keep the prompt prefix identical, and the repeated part is billed at the $0.0125 cached rate.
We used a support handbook of about 3,290 tokens as the system prompt and asked four different questions in a row. The first three missed the cache at about $0.001 each. The fourth hit 3,072 cached tokens and cost $0.00013. For support bots and fixed templates with long, repeated prefixes, caching is the biggest saving available.
How it compares on HiAPI
| Model | Input / M | Output / M | One order extraction (250 in / 200 out) |
|---|---|---|---|
| GPT-6.1 Terra | $0.25 | $1.25 | $0.0003 |
| GPT-6.1 Sol | $0.50 | $2.50 | $0.0006 |
| GPT-5.6 Terra | $0.50 | $3.00 | $0.0007 |
| GPT-5.6 Luna | $0.05 | $0.30 | $0.00007 |
| DeepSeek V4.1 Flash | $0.44 | $1.32 | $0.0004 |
| Claude Opus 5.5 | $2.02 | $10.08 | $0.0025 |
| Claude Sonnet 4.6 | $3.00 | $15.00 | $0.0038 |
All HiAPI prices, per million tokens.
The previous GPT-5.6 Terra is $0.50 / $3.00, so 6.1 Terra halves input and cuts output by about 58%. It's also half the price of GPT-6.1 Sol. For a same-task comparison, see GPT-6.1 Terra vs GPT-6.1 Sol.
Monthly budget
| Task (tokens per call) | 1,000 calls | 10,000 calls | 100,000 calls |
|---|---|---|---|
| Extraction (250 in / 200 out) | $0.31 | $3.13 | $31.25 |
| Code review (150 in / 2,500 out) | $3.16 | $31.63 | $316.25 |
| Long-document summary (100K in / 1,500 out) | $26.88 | $268.75 | $2,687.50 |
| Very long document (300K in / 2,000 out, long-context rate) | $153.75 | $1,537.50 | $15,375 |
Monthly cost = calls × tokens per call × rate. Output tokens cost five times as much as input, so long answers add up. For long documents the input dominates, so cache any prefix you can.
FAQ
How much cheaper is GPT-6.1 Terra on HiAPI than OpenAI's list price?
75% cheaper. OpenAI lists $1 input and $5 output; HiAPI charges $0.25 and $1.25. The long-context rates above 272K are also a quarter of OpenAI's.
Are reasoning tokens billed separately?
No. They're already included in output_tokens and billed once at the output rate.
Is there a Batch discount?
Every HiAPI request uses the standard rates above, which are already below OpenAI's Batch price of $0.50 / $2.50. Run offline jobs through the normal API.
Am I charged for failed requests?
No. When the stream returns response.failed, usage is empty and nothing is billed. Retry the request.
Integration code is in the GPT-6.1 Terra API guide, and ready-made prompts are in GPT-6.1 Terra prompts. Try one of your own tasks first: open the GPT-6.1 Terra model page and check the token usage and estimated cost shown when it finishes.
