Cheapest LLM per solved coding task
- models qualify
- 84
Cost per solved task · read October 20
Cheapest output, $ per million tokens
- DeepSeek V4.1 Flash $0.0112
- Mistral Nemo $0.03
- Gemma 3 4B $0.05
- S Llama 3 8B Lunaris $0.05
- Ling 3.0 Flash $0.063
- Mistral Small 3 $0.08
- Llama 3.1 8B Instruct $0.08
- gpt-oss-20b $0.09
As of October 20, 2026, DeepSeek V4 Flash 0731 solves a SWE-bench Verified task for $0.01, the lowest of the 84 models where that cost can be computed, ahead of Xiaomi MiMo-V2.5 ($0.01) and Z.ai GLM 5.3 Flash ($0.02).
This list narrows the best LLM for coding ranking.
| Rank | Model | Cost per solved task | Evidence | SWE-bench resolved | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4 Flash 0731 DeepSeek | $0.01 | 88.8% | Not benchmarked | $1.32 | |
| 2 | MiMo-V2.5 Xiaomi | $0.01 | 71.0% | 34.0 | $0.28 | |
| 3 | GLM 5.3 Flash Z.ai | $0.02 | 92.0% | 71.0 | $0.16 | |
| 4 | MiMo-V2.5-Pro Xiaomi | $0.03 | 74.0% | 40.0 | $0.87 | |
| 5 | GPT-5.6 Luna OpenAI | $0.04 | 93.0% | 61.5 | $1.20 | |
| 6 | Ling 3.0 Flash inclusionAI | $0.05 | 65.2% | Not benchmarked | $0.063 | |
| 7 | Nemotron 3.5 Lightning NVIDIA | $0.06 | 52.8% | Not benchmarked | $0.20 | |
| 8 | GPT-5 Mini OpenAI | $0.08 | 60.8% | 59.8 | $2.00 | |
| 9 | DeepSeek V4 Pro 0813 DeepSeek | $0.10 | 96.4% | 54.6 | $1.98 | |
| 10 | GPT-5 Nano OpenAI | $0.109 | 34.8% | 34.8 | $0.40 | |
| 11 | GPT-5.4 Nano OpenAI | $0.14 | 69.8% | Not benchmarked | $1.25 | |
| 12 | T Inkling Small Thinking Machines | $0.16 | 82.2% | 29.9 | $1.20 | |
| 13 | $0.18 | 62.8% | 8.1 | $1.50 | ||
| 14 | Kimi K2.5 Moonshot AI | $0.24 | 70.0% | 58.5 | $2.50 | |
| 15 | GLM 4.7 Z.ai | $0.29 | 69.4% | 33.6 | $1.98 | |
| 16 | Grok 4.3 xAI | $0.31 | 71.4% | 22.5 | $2.50 | |
| 17 | Kimi K2.6 Moonshot AI | $0.35 | 76.2% | 44.2 | $3.41 | |
| 18 | Kimi K2.7 Code Moonshot AI | $0.35 | 78.2% | 39.1 | $3.35 | |
| 19 | GLM 5.3 Z.ai | $0.36 | 95.4% | 73.1 | $1.40 | |
| 20 | GPT-4.1 OpenAI | $0.37 | 39.6% | 46.5 | $8.00 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Every model with a cost per solved SWE-bench Verified task: what one attempt costs, divided by the share of tasks the model resolves. Cheapest first. A model without a published attempt cost is not on the list.
- OpenRouter, fetched
- SWE-bench (swebench.com), fetched
- Vals AI boards, fetched