BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases15

LiveBench 2.0 Coding

LiveBench 2.0 coding suite: real merged-PR bug fixes on a frozen monorepo, graded deterministically.

As of 21 Oct 2026, Claude Fable 5 leads LiveBench 2.0 Coding on BenchLeader with 98.3%, ahead of Claude Fable 5.1 at 97.7%, across 48 model configurations with a published result.

Published by
LiveBench
Category
Coding
Index weight
Reference only
Models
48
Data as of
21 Oct 2026

LiveBench (livebench.ai), White et al. 2024.

Not comparable with the earlier test of the same name, whose final scores stay on LiveBench 1.0 Coding.

What the test looks like

LiveBench 2.0's coding suite: the model fixes real bugs from merged pull requests in a frozen monorepo, working with tools in an agent loop.

How it is scored

Mean task score from 0 to 100 on the current release, graded deterministically; a fix that breaks something else scores 0.

What to keep in mind

It shares a name with LiveBench 1.0's Coding category, which was short competitive-programming problems, so the two are kept apart; 2.0 scores start on 19 October 2026. Harness-dependent, like any agentic test.

Top 15
  1. 1Claude Fable 598.3%
  2. 2Claude Fable 5.197.7%
  3. 3Claude Opus 5.596.3%
  4. 4GPT-6.1 Sol96.0%
  5. 5DeepSeek V4.1 Flash 101796.0%
  6. 6Claude Sonnet 5.594.7%
  7. 7Grok 4.594.3%
  8. 8GPT-6 Astra93.7%
  9. 9GPT-6 Sol93.1%
  10. 10Claude Haiku 5.591.4%
  11. 11Claude Opus 590.8%
  12. 12Muse Spark 1.390.2%
  13. 13DeepSeek V4.1 Flash90.2%
  14. 14Gemini 4 Argon89.7%
  15. 15GPT-5.6 Sol89.7%
48 of 48
#
1Claude Fable 598.3%67.12026-10-14
2Claude Fable 5.197.7%66.22026-10-14
3Claude Opus 5.596.3%61.92026-10-14
4GPT-6.1 Sol96.0%63.32026-10-14
5NewDeepSeek V4.1 Flash 101796.0%–2026-10-14
6Claude Sonnet 5.594.7%–2026-10-14
7Grok 4.594.3%58.92026-10-14
8GPT-6 Astra93.7%67.92026-10-14
9GPT-6 Sol93.1%–2026-10-14
10Claude Haiku 5.591.4%–2026-10-14
11Claude Opus 590.8%65.42026-10-14
12Muse Spark 1.390.2%58.32026-10-14
13DeepSeek V4.1 Flash90.2%–2026-10-14
14Gemini 4 Argon89.7%63.32026-10-14
15GPT-5.6 Sol89.7%55.92026-10-14
16Grok 4.689.1%57.82026-10-14
17Qwen3.8 Flash87.4%–2026-10-14
18Qwen3.8 27B86.8%54.52026-10-14
19Muse Wonder86.2%63.62026-10-14
20GPT-5.6 Terra86.2%52.62026-10-14
21Step 5 Preview85.6%60.62026-10-14
22GPT-5.585.1%60.22026-10-14
23Grok 4.784.5%–2026-10-14
24GPT-6 Luna84.5%–2026-10-14
25NewMiniMax-M3.1-Flash83.9%61.92026-10-14
26NewBeam 1.083.9%54.62026-10-14
27GLM 5.382.8%–2026-10-14
28Gemini 3.8 Flash82.2%55.62026-10-14
29MiMo-V2.6-Pro81.5%64.12026-10-14
30GLM 5.3 Flash80.6%62.72026-10-14
31Gemini 3.7 Flash80.5%53.22026-10-14
32Muse Spark 1.278.2%–2026-10-14
33Claude Opus 4.877.3%60.12026-10-14
34Claude Sonnet 574.7%56.02026-10-14
35GPT-5.6 Luna73.6%47.92026-10-14
36NewMuse Spark Open73.6%–2026-10-14
37MiniMax M370.9%56.52026-10-14
38Gemini 3.1 Pro66.7%60.72026-10-14
39Claude Sonnet 4.666.6%57.12026-10-14
40Nemotron Super VL 3.565.5%–2026-10-14
41Qwen3.8 Maxmax60.3%63.32026-10-14
42Mistral Large 460.0%56.72026-10-14
43Kimi K359.3%59.12026-10-14
44Gemma 4 31B49.0%48.32026-10-14
45Claude Haiku 4.548.1%45.32026-10-14
46Nemotron Ultra 343.6%–2026-10-14
47Deepseek v4 Flash Vision37.6%–2026-10-14
48Nemotron Lightning 3.522.6%–2026-10-14

More coding benchmarks · All LLM benchmarks

Cite as: BenchLeader, “LiveBench 2.0 Coding leaderboard”, https://www.benchleader.com/benchmarks/livebench2_coding, data as of 21 Oct 2026.

LiveBench 2.0 Coding: questions

What does LiveBench 2.0 Coding measure?
LiveBench 2.0's coding suite: the model fixes real bugs from merged pull requests in a frozen monorepo, working with tools in an agent loop. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads LiveBench 2.0 Coding?
Claude Fable 5 leads LiveBench 2.0 Coding with 98.3% as of 21 Oct 2026, ahead of Claude Fable 5.1 at 97.7%.
How many models have LiveBench 2.0 Coding results?
48 model configurations have a LiveBench 2.0 Coding result on BenchLeader, all taken from LiveBench.
Who runs LiveBench 2.0 Coding and how often is it updated?
LiveBench 2.0 Coding is published by LiveBench. BenchLeader re-reads the published results once a day and records the date each result was published.
Does LiveBench 2.0 Coding count toward the BenchLeader Index?
No. LiveBench 2.0 Coding is shown for reference but left out of the composite index.