TL;DR
- GPT-6 Luna (
gpt-6-luna) on HiAPI charges per token: $0.025/M input, $0.125/M output, $0.0025/M cached read for prompts up to 272K tokens; long-context (>272K) is $0.05/M input and $0.1875/M output. - OpenAI lists it at $0.10/$0.50. HiAPI's price is about a quarter of that, making Luna the cheapest GPT text model on the platform.
- GPT-6 Luna is live on HiAPI. Open the model page to chat directly, and new accounts get 200 free credits on sign-up.
How much does GPT-6 Luna cost?
| Tier | Input ($/M token) | Output ($/M token) | Cached read ($/M token) |
|---|---|---|---|
| Standard (≤272K) | $0.025 | $0.125 | $0.0025 |
| Long context (>272K) | $0.05 | $0.1875 | — |
You pay actual tokens consumed at the corresponding rate. Cached reads apply when you resend the same context: the cached portion is billed at the much lower rate. The long-context tier kicks in automatically when a single request exceeds 272K input tokens.
What a first test costs: $1 handles about 40M input tokens or 8M output tokens at the standard tier. Run a thousand articles through classification and you will still have budget to spare. Try your first prompt on the model page →
How much cheaper is it than the list price?
| HiAPI (standard tier) | OpenAI list price | HiAPI as % of list | |
|---|---|---|---|
| Input | $0.025/M | $0.10/M | 25% |
| Output | $0.125/M | $0.50/M | 25% |
| Cached read | $0.0025/M | $0.025/M | 10% |
HiAPI's price is roughly a quarter of the list price, with cached reads dropping to a tenth.
Four ways to spend even less
- Cache hits: $0.0025/M. System prompts, few-shot examples and repeated context segments are nearly free. Caching is automatic and needs no configuration.
- Batch or Flex processing: 50% off. Classification, extraction and rewriting jobs that do not need real-time responses can run through Batch or Flex at $0.0125/M input and $0.0625/M output.
- Turn off reasoning:
reasoning.effort=none. Luna defaults tomediumeffort. Setting it tononelets you freely adjust temperature and top_p, and stops reasoning-token consumption entirely. Tasks like classification, extraction and structured output rarely need reasoning overhead. - Tune prompts at small scale first. At $0.025/M input, iterating on prompts costs virtually nothing.
How Luna stacks up against other GPT models on HiAPI
| Model | Input ($/M) | Output ($/M) | Cached read ($/M) | vs Luna |
|---|---|---|---|---|
| gpt-6-luna | $0.025 | $0.125 | $0.0025 | — |
| gpt-5.6-luna | $0.05 | $0.30 | $0.005 | Half the input, 2.4x cheaper output |
| gpt-6.1-terra | $0.25 | $1.25 | $0.0125 | 10x Luna on input |
| gpt-6.1-sol | $0.50 | $2.50 | $0.025 | 20x Luna on input |
Luna is the most affordable GPT text model on HiAPI. Terra and Sol offer stronger reasoning and larger context windows: pick them for complex multi-step analysis. For batch classification, extraction, rewriting and simple agent loops, Luna is the right tool at the right price.
Monthly budget table
| Monthly volume | Input only | Output only | Input:Output = 4:1 (typical) | Input:Output = 1:1 (chat-heavy) |
|---|---|---|---|---|
| 1M tokens | $0.025 | $0.125 | $0.045 | $0.075 |
| 10M tokens | $0.25 | $1.25 | $0.45 | $0.75 |
| 100M tokens | $2.50 | $12.50 | $4.50 | $7.50 |
| 1B tokens | $25 | $125 | $45 | $75 |
Actual cost = input tokens x $0.025 + output tokens x $0.125 minus cached-read savings. Typical batch workloads are input-heavy: 1B tokens a month runs about $45.
FAQ
What is GPT-6 Luna best for? Batch text classification, structured data extraction, content rewriting, automated FAQ replies, and low-complexity agent loops. For deep multi-step reasoning, reach for Terra or Sol instead.
How does caching work? Cache is automatic. Under the same API key, repeated system prompts and message history segments hit the cache and are billed at $0.0025/M. No additional setup required.
How do I use Batch or Flex?
Add "processing_mode": "batch" or "processing_mode": "flex" to your request. Tasks are queued asynchronously at half price. Best for workloads that do not need instant responses.
What if my output exceeds 128K tokens? Luna caps output at 128K tokens. For longer documents, request an outline first, then process in sections and merge. This pattern works well for summarisation and rewriting at scale.
How hard is it to switch from gpt-5.6-luna?
Trivial. Both models support the Chat Completions protocol: change the model field and you are done. To use new features like the Responses API or six effort levels, you will need to adapt the request format, but standard chat calls work as-is. See the GPT-6 Luna vs GPT-5.6 Luna comparison.
Ready to try? Open the model page and start chatting. The 200 sign-up credits cover hundreds of rounds. Or read the API guide to fire off your first call with curl.
