LLM prices / Cheapest
AI API prices · 29 current models · checked Oct 23, 2026
Cheapest LLM API for every job
Short answerIt depends on the job, and right now it is a tie: Claude Haiku 5.5 and GPT-6 Luna tie for cheapest on all 8 of our jobs because their list prices are identical. Prices checked Oct 23, 2026; current models only; ranked by cost, not quality.
| Job | Cheapest | Per 1,000 | Runner-up |
|---|---|---|---|
| Chatbota 1,500-token prompt (system prompt plus the conversation so far) and a 400-token reply... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.35 | GPT-5.6 Luna $0.78 |
| RAG (answer from your docs)6,000 tokens of retrieved passages plus the question, and a 500-token answer... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.85 | GPT-5.6 Luna $1.80 |
| Coding agent40,000 tokens of repo context and history, 30,000 of them repeated from the last step and ... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $2.30 | GPT-5.6 Luna $5.00 |
| Tool-using agent12,000 tokens of instructions, tool definitions and history, 10,000 of them repeated from ... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.60 | GPT-5.6 Luna $1.32 |
| Vision (read an image)one image counted as 1,000 input tokens plus a 300-token question, and a 300-token answer.... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.28 | GPT-5.6 Luna $0.62 |
| Translationa 1,000-token passage in and about 1,100 tokens out (translations run a little longer than... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.65 | GPT-5.6 Luna $1.52 |
| Batch extraction2,000 tokens per record in and 300 tokens of JSON out, through the batch API where there i... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $0.17 | GPT-5.6 Luna $0.38 |
| Summarize long documentsa 20,000-token document in and an 800-token summary out... | Claude Haiku 5.5 GPT-6 Lunatie: same price | $2.40 | GPT-5.6 Luna $4.96 |
Each job is a fixed request size, written out on its page. Billed the cheapest way each price list allows (cache-read prices for repeated context, batch where the job can wait). Cache writes not counted.