Best LLM for competitive programming
- Claude Fable 5.1 leads
- 90.5%
- models ranked
- 95
- lead over #2
- +0.7 pts
LiveCodeBench (Vals AI) · read October 21
# LiveCodeBench (Vals AI), read October 21
leaders = {
"Claude Fable 5.1": 90.5%, "Claude Fable 5": 89.8%, "Gemini 3.8 Flash": 89.5%,
}
As of October 21, 2026, Anthropic Claude Fable 5.1 leads 95 models on LiveCodeBench (Vals AI) with 90.5%, ahead of Anthropic Claude Fable 5 (89.8%) and Google Gemini 3.8 Flash (89.5%).
The full best LLM for coding ranking combines this benchmark with others.
| Rank | Model | LiveCodeBench (Vals AI) | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.1 Anthropic | 90.5% | 84.2 | $50.00 | |
| 2 | Claude Fable 5 Anthropic | 89.8% | 79.3 | $50.00 | |
| 3 | Gemini 3.8 Flash Google | 89.5% | 61.6 | $3.75 | |
| 4 | Claude Opus 5 Anthropic | 89.0% | 83.0 | $25.00 | |
| 5 | Gemini 3.7 Flash Google | 88.7% | 58.0 | $3.75 | |
| 6 | Gemini 3.1 Pro Preview Google | 88.5% | 32.3 | $12.00 | |
| 7 | Grok 4.6 xAI | 88.2% | 65.7 | $6.00 | |
| 8 | Gemini 3.6 Flash Google | 88.1% | 62.5 | $3.75 | |
| 9 | GPT-5.2-Codex OpenAI | 88.0% | 45.4 | $14.00 | |
| 10 | Qwen3.8 Max (0902) Qwen | 87.8% | 62.4 | $6.00 | |
| 11 | Claude Opus 4.8 Anthropic | 87.8% | 65.0 | $25.00 | |
| 12 | Gemini 3.5 Flash Google | 87.6% | 39.8 | $9.00 | |
| 13 | DeepSeek V4 Pro 0813 DeepSeek | 87.5% | 56.1 | $1.98 | |
| 14 | DeepSeek V4 Pro 0423 DeepSeek | 87.5% | 38.7 | $2.716 | |
| 15 | Grok 4.5 xAI | 87.3% | 58.6 | $6.00 | |
| 16 | GPT-5.3-Codex OpenAI | 87.3% | 60.0 | $14.00 | |
| 17 | DeepSeek V4 Flash 0731 DeepSeek | 87.3% | 51.8 | $0.28 | |
| 18 | Kimi K3 Moonshot AI | 87.2% | 62.3 | $15.00 | |
| 19 | Qwen3.7 Max Qwen | 87.1% | 54.1 | $4.425 | |
| 20 | Kimi K2.6 Moonshot AI | 86.8% | 52.1 | $4.00 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Models by their LiveCodeBench (Vals AI) result, as the board publishes it, best first. Competitive programming problems collected after the models were trained, run by Vals AI. Share solved.
- OpenRouter, fetched
- Vals AI boards, fetched