LLM prices / Cheapest / Tool-using agent
AI API prices · 29 models fit · checked Oct 23, 2026
Cheapest LLM for agents
Short answerClaude Haiku 5.5 and GPT-6 Luna tie for the cheapest current model for a agent step at $0.60 per 1,000, with the repeated 10,000 tokens billed at the cache-read price. At 10,000 a month that is $6.00. Next: GPT-5.6 Luna at $1.32 and DeepSeek Flash at $1.38 per 1,000. 29 of 29 current models fit this job. Ranked by cost only; we don't score quality.
Every model that fits, cheapest first
One agent step = 12,000 tokens of instructions, tool definitions and history, 10,000 of them repeated from the last step and cached, 600 tokens out, tool calling required. Each model is billed the cheapest way its price list allows.
| # | Model | Per 1,000 | 10,000 a month | 100,000 a month | Billed | Context |
|---|---|---|---|---|---|---|
| 1 | Claude Haiku 5.5Anthropic | $0.60 | $6.00 | $60.00 | cached | 1M |
| 2 | GPT-6 LunaOpenAI | $0.60 | $6.00 | $60.00 | cached | 1.05M |
| 3 | GPT-5.6 LunaOpenAI | $1.32 | $13.20 | $132 | cached | 1.05M |
| 4 | DeepSeek FlashDeepSeek | $1.38 | $13.80 | $138 | cached | 1M |
| 5 | Gemini 3.1 Flash-LiteGoogle | $1.65 | $16.50 | $165 | cached | 1.05M |
| 6 | Gemini 3.5 Flash-LiteGoogle | $2.40 | $24.00 | $240 | cached | 1.05M |
| 7 | Gemini 3 Flash PreviewpreviewGoogle | $3.30 | $33.00 | $330 | cached | 1.05M |
| 8 | Gemini 3.8 FlashGoogle | $4.50 | $45.00 | $450 | cached | 1.05M |
| 9 | Gemini 3.6 FlashGoogle | $4.50 | $45.00 | $450 | cached | 1.05M |
| 10 | GPT-5.4 miniOpenAI | $4.95 | $49.50 | $495 | cached | 400K |
| 11 | DeepSeek V4 ProDeepSeek | $5.46 | $54.56 | $546 | cached | 1M |
| 12 | Claude Sonnet 5.5Anthropic | $11.00 | $110 | $1,100 | cached | 1M |
| 13 | GPT-6.1 SolOpenAI | $11.00 | $110 | $1,100 | cached | 1.05M |
| 14 | Gemini 4 ArgonGoogle | $11.00 | $110 | $1,100 | cached | 1.05M |
| 15 | GPT-6 SolOpenAI | $12.00 | $120 | $1,200 | cached | 1.05M |
| 16 | GPT-5.6 TerraOpenAI | $13.20 | $132 | $1,320 | cached | 1.05M |
| 17 | Gemini 3.1 Pro PreviewpreviewGoogle | $13.20 | $132 | $1,320 | cached | 1.05M |
| 18 | GPT-5.2OpenAI | $13.65 | $136 | $1,365 | cached | 400K |
| 19 | GPT-5.4OpenAI | $16.50 | $165 | $1,650 | cached | 1.05M |
| 20 | Claude Opus 5.5Anthropic | $22.00 | $220 | $2,200 | cached | 1M |
| 21 | GPT-5.6 SolOpenAI | $24.00 | $240 | $2,400 | cached | 1.05M |
| 22 | GPT-5.5OpenAI | $33.00 | $330 | $3,300 | cached | 1.05M |
| 23 | chat-latestOpenAI | $33.00 | $330 | $3,300 | cached | 400K |
| 24 | Claude Fable 5.1Anthropic | $52.50 | $525 | $5,250 | cached | 1M |
| 25 | GPT-6 AstraOpenAI | $60.00 | $600 | $6,000 | cached | 1.05M |
| 26 | GPT-5.6 CyberOpenAI | $82.50 | $825 | $8,250 | cached | 400K |
| 27 | GPT-5.2 ProOpenAI | $353 | $3,528 | $35,280 | standard | 400K |
| 28 | GPT-5.5 ProOpenAI | $468 | $4,680 | $46,800 | standard | 1.05M |
| 29 | GPT-5.4 ProOpenAI | $468 | $4,680 | $46,800 | standard | 1.05M |
What moves the answer
- Caching decides it. 10,000 of the 12,000 input tokens repeat every step, so cache-read prices matter more than list prices: 10 of the top 10 are cheapest with caching on. Without caching the winner would be Claude Haiku 5.5 at $1.50 per 1,000.
- Input price matters most here. The job reads 20x more than it writes, so the input (and cache-read) price drives the bill.
- Watch long prompts. 13 of these models charge more once a prompt passes a size threshold (for example Claude Haiku 5.5 above 100,000 tokens); this job stays under it, a bigger one may not.
Method
Prices: each provider's own pricing page, checked Oct 23, 2026. Fits: current on the provider's model page (not deprecated, legacy or shut down), lists tool calling, and a context window that holds the request; checked Oct 24, 2026. Cache writes are not counted. This page ranks price only; quality on your task is a separate test. People search for this: Google suggests "cheapest llm for agents".