Which AI model is cheapest per coding-agent request?

GPT-6 Luna is the cheapest current model per coding-agent request at list prices, about $0.0041 per average request, followed by DeepSeek V4.1 Flash at $0.0067 and Claude Haiku 5.5 at $0.013.

List prices: OpenAI API pricing, DeepSeek API pricing, Mistral API pricing, OpenRouter (vendor endpoint), Anthropic pricing, Z.AI pricing, Gemini API pricing, xAI models and pricing, checked 2026-10-13. Request costs: Standard Compute gateway data, 79,301 coding-agent requests from 75 users, Sep 13 to Oct 12, 2026. How we compute this

Run 17 of these 32 models in your agent from $19/month

See plans

24 models ranked by cost per request

Model	List price in / out	Average request	100 requests
1. GPT-6 Luna11% of requests at its long-prompt rate	$0.10 / $0.50	$0.0041	$0.41
2. DeepSeek V4.1 Flash	$0.30 / $1.20	$0.0067	$0.67
3. GPT-5.6 Lunasuperseded, 11% of requests at its long-prompt rate	$0.20 / $1.20	$0.0084	$0.84
4. Claude Haiku 5.550% of requests at its long-prompt rate	$0.10 / $0.50	$0.013	$1.31
5. DeepSeek V4 Pro	$1.32 / $3.96	$0.030	$3.04
6. Gemini 3.8 Flash	$1.50 / $7.50	$0.048	$4.83
7. Claude Sonnet 5.5	$2 / $10	$0.053	$5.33
8. GLM 5.3	$1.40 / $4.40	$0.057	$5.72
9. GLM 5.2superseded	$1.40 / $4.40	$0.057	$5.72
10. GPT-6.1 Sol11% of requests at its long-prompt rate	$2 / $10	$0.067	$6.74
11. Grok 4.2021% of requests at its long-prompt rate	$1.25 / $2.50	$0.068	$6.84
12. Grok 4.3superseded, 21% of requests at its long-prompt rate	$1.25 / $2.50	$0.068	$6.84
13. GPT-6 Solsuperseded, 11% of requests at its long-prompt rate	$2 / $10	$0.083	$8.25
14. GPT-5.6 Terra11% of requests at its long-prompt rate	$2 / $12	$0.084	$8.40
15. Gemini 3.1 Pro21% of requests at its long-prompt rate	$2 / $12	$0.095	$9.48
16. Claude Sonnet 4.6superseded	$3 / $15	$0.097	$9.74
17. GPT-5.411% of requests at its long-prompt rate	$2.50 / $15	$0.10	$10.49
18. Claude Opus 5.5	$4 / $20	$0.11	$10.66
19. Claude Opus 4.8superseded	$5 / $25	$0.16	$16.23
20. GPT-5.6 Solsuperseded, 11% of requests at its long-prompt rate	$4 / $20	$0.17	$16.50
21. GPT-5.5superseded, 11% of requests at its long-prompt rate	$5 / $30	$0.21	$20.99
22. Claude Fable 5.1	$10 / $50	$0.24	$23.76
23. Claude Fable 5superseded	$10 / $50	$0.32	$32.47
24. GPT-6 Astra11% of requests at its long-prompt rate	$10 / $50	$0.41	$41.25

Every one of 79,301 coding-agent requests (Sep 13 to Oct 12, 2026) priced at each vendor's list rates for its own token counts, cache reads and any higher rate for long prompts. Methodology

Models whose context window is too small for some requests

These cannot take the largest prompts in our sample, so their figures cover only the requests that fit and are not ranked with the others.

Model	List price in / out	Average request	100 requests
Codestral13% of requests too large	$0.30 / $0.90	$0.0070	$0.70
Devstral 212% of requests too large	$0.40 / $2	$0.010	$1.01
Mistral Large 312% of requests too large	$0.50 / $1.50	$0.012	$1.19
GLM 4.720% of requests too large	$0.60 / $2.20	$0.017	$1.68
Claude Haiku 4.5superseded, 21% of requests too large	$1 / $5	$0.022	$2.22
GPT-5.4 Mini3% of requests too large	$0.75 / $4.50	$0.023	$2.27
GPT-5.3-Codex3% of requests too large	$1.75 / $14	$0.055	$5.55
Grok 4.719% of requests at its long-prompt rate, 1% of requests too large	$2 / $6	$0.13	$13.16

What to watch for

This ranks price. It does not measure ability: the cheapest models are small and weaker on long agent tasks, so they can need more steps for the same result, and prices for DeepSeek are its peak-hour rates (half price off-peak). Cost per finished task depends on your work; test two or three candidates on a task from your own repository.

FAQ

Which AI model is cheapest per coding-agent request?

At list prices, per 100 average coding-agent requests: GPT-6 Luna $0.41, DeepSeek V4.1 Flash $0.67, Claude Haiku 5.5 $1.31, DeepSeek V4 Pro $3.04, Gemini 3.8 Flash $4.83. Each request in our sample was priced with its own token counts and cache reads.

Why does the ranking differ from price per million tokens?

Coding agents resend their whole context on every step, so cache-read prices and long-context surcharges decide the bill. Claude Haiku 5.5 and GPT-6 Luna list at the same $0.10/$0.50, but 50% of requests in our data passed Haiku 5.5's 100K line, where every token costs five times more, so Haiku 5.5 costs $0.013 per request against Luna's $0.0041.

What does a typical coding-agent request look like?

In our data the median request sent 99,534 input tokens (p90 279,918), 87% of input tokens were cache reads, and requests returned 698 output tokens on average. A median session ran 16 requests.

Is the cheapest model per request the cheapest per finished task?

Not necessarily. A weaker model can take more steps or retries to finish the same task, which this ranking does not measure. Use it to compare like with like, then test the two or three candidates on a task from your own repository.

Current Claude, GPT, Gemini and open models are included with Standard Compute

Standard Compute

Current frontier models Claude Opus 5.5 · Claude Sonnet 5.5