The ChatGPT API costs a fraction of a cent per message on the smaller OpenAI models. In my measurements, GPT-6 Luna answers a chat question for 0.06¢ per task, or $0.59 per 1,000. GPT-6 Astra costs 3.4¢ for the same task, or $33.60 per 1,000.
What a "message" means here
The API bills per token, not per message. So I run a fixed task and record the tokens used. This article covers three tasks:
- Chat answer: a single question answered in one reply.
- Email reply: answering a customer email.
- RAG answer: answering a question from five retrieved documents, which means a much longer input.
Reasoning tokens are billed as output tokens, so a model that thinks before it replies pays for that thinking. Costs below are per task, with the per 1,000 figure alongside. See the methodology for how the runs were set up.
OpenAI list prices
Prices per 1M tokens (input / output / cached input):
| Model | Input | Output | Cached input | Reasoning |
|---|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | default |
| GPT-5.4 Nano | $0.20 | $1.25 | $0.02 | default |
| GPT-5.4 Mini | $0.75 | $4.50 | $0.075 | default |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 | always |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 | always |
Price per token is only part of the story. What matters is how many tokens a task uses, and reasoning models can use many more output tokens than they show.
Cost per chat answer
| Model | Per chat answer | Per 1,000 | Basis |
|---|---|---|---|
| GPT-6 Luna | 0.06¢ | $0.59 | measured |
| GPT-5.4 Nano | 0.15¢ | $1.48 | estimated |
| GPT-5.4 Mini | 0.53¢ | $5.34 | estimated |
| GPT-6.1 Sol | 0.67¢ | $6.72 | measured |
| GPT-6 Astra | 3.4¢ | $33.60 | estimated |
The median across all 326 models is 0.24¢ per chat answer. Luna is well under it, and Astra is well over it.
Luna's run used 45 input, 540 output and 639 reasoning tokens. Sol used 45 input, 560 output and 103 reasoning tokens. Astra's estimate assumes the same 45 / 560 / 103 profile as Sol, so its higher cost comes from its $10.00 input and $50.00 output list prices.
Other labs on the same chat task:
- Claude Haiku 5.5: 0.07¢ ($0.66 per 1,000), measured
- Claude Sonnet 5.5: 1.3¢ ($12.68 per 1,000), measured
- Claude Opus 5.5: 2.5¢ ($25.36 per 1,000), estimated
- Claude Fable 5.1: 6.3¢ ($63.41 per 1,000), estimated
Anthropic's tokenizer produces 1.39 times as many tokens as OpenAI's for the same English document, which affects how list prices translate into per-task cost.
Cost per email reply
| Model | Per email reply | Per 1,000 | Basis |
|---|---|---|---|
| GPT-6 Luna | under 0.01¢ | $0.07 | measured |
| GPT-5.4 Nano | 0.02¢ | $0.17 | estimated |
| GPT-5.4 Mini | 0.06¢ | $0.62 | estimated |
| GPT-6.1 Sol | 0.18¢ | $1.78 | measured |
| GPT-6 Astra | 0.89¢ | $8.88 | estimated |
Email replies are short, so most models are cheap. The median across all models is 0.10¢. Luna, Nano and Mini all sit below it, while Sol and Astra sit above it.
Sol's measured run used 258 input, 126 output and 0 reasoning tokens, so its cost tracks the list price closely. At scale, the gap between Luna and Astra is $0.07 versus $8.88 per 1,000 replies.
Cost per RAG answer
| Model | Per RAG answer | Per 1,000 | Basis |
|---|---|---|---|
| GPT-6 Luna | 0.05¢ | $0.54 | measured |
| GPT-5.4 Nano | 0.12¢ | $1.16 | estimated |
| GPT-5.4 Mini | 0.43¢ | $4.29 | estimated |
| GPT-6.1 Sol | 1.2¢ | $11.80 | measured |
| GPT-6 Astra | 5.9¢ | $59.00 | estimated |
RAG is where input price starts to matter more. The OpenAI runs send 3,690 input tokens across the five documents, so the input rate counts for much more than in a short chat question. The median across all models is 0.50¢.
Luna's measured run used 3,690 input, 234 output and 104 reasoning tokens. Sol's used 3,690 input, 442 output and no reasoning tokens.
Other labs on the RAG task:
- Claude Haiku 5.5: 0.14¢ ($1.44 per 1,000), measured
- Gemini 3.8 Flash: 1.1¢ ($10.84 per 1,000), measured
- Claude Opus 5.5: 5.5¢ ($54.62 per 1,000), estimated
Live RAG ranking:
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GPT-6 LunaOpenAI | 0.05¢$0.54 / 1,000 | 1443 | 3,690 / 338 | measured best value | |
| 2 |  GLM 5.3 FlashZ.ai | 0.06¢$0.57 / 1,000 | 1474 | 3,714 / 1,971 | measured best value | |
| 3 |  DeepSeek V4.1 FlashDeepSeek | 0.08¢$0.81 / 1,000 | 1475 | 3,696 / 3,750 | measured best value | |
| 4 | GPT-5.4 NanoOpenAI | 0.12¢$1.16 / 1,000 | 1401 | 3,690 / 338 | estimated | |
| 5 |  Claude Haiku 5.5Anthropic | 0.14¢$1.44 / 1,000 | - | 5,184 / 1,849 | measured | |
| 6 |  Qwen3.8 FlashAlibaba Qwen | 0.19¢$1.92 / 1,000 | - | 3,751 / 2,886 | measured | |
| 7 |  Mistral Small 4Mistral | 0.25¢$2.48 / 1,000 | - | 3,772 / 3,191 | estimated | |
| 8 |  Llama 4 MaverickMeta | 0.26¢$2.58 / 1,000 | 1327 | 3,721 / 2,878 | estimated | |
| 9 |  MiniMax M3MiniMax | 0.26¢$2.64 / 1,000 | 1441 | 3,773 / 1,259 | measured | |
| 10 |  MiMo-V2.6-ProXiaomi | 0.28¢$2.76 / 1,000 | 1479 | 3,807 / 1,266 | estimated best value |
How to read these numbers
- Pick by task, not by brand. GPT-6 Luna is the cheapest OpenAI option on all three tasks in my data. For simple jobs you may not need Astra.
- Check input versus output. Chat is short, while RAG sends thousands of input tokens. The ranking can shift with the shape of your task.
- Know what is measured. GPT-6 Luna and GPT-6.1 Sol were measured. GPT-5.4 Nano, GPT-5.4 Mini and GPT-6 Astra are estimated. Use estimates for ranking, not as invoices.
- Compare per-task cost, not list price. A cheap model that reasons at length can cost more than a pricier one that doesn't.
Use the calculator to plug in your own token counts. Model pages: GPT-6 Luna, GPT-6.1 Sol and GPT-6 Astra. Task pages: chat answer, email reply and RAG answer.
FAQ
Is the ChatGPT API priced per message?
No. The API bills per token, both input and output, and reasoning tokens count as output. A message's cost depends on prompt length and how much the model reasons before answering.
Which OpenAI model is cheapest per chat answer?
GPT-6 Luna, at 0.06¢ per chat answer or $0.59 per 1,000. GPT-5.4 Nano is next at 0.15¢, though that figure is estimated.
Are these prices what I will pay?
They are list prices as of 2026-10-11 and per-task costs from my runs. Measured models had one run per task with default settings. Your bill will vary with prompt length, output length and caching.