BenchLeader
Menu
LeaderboardsModelsProvidersBenchmarksCompareReleases1BlogMethodology

Vibe Code Bench 1-100

Extending a working web app, request by request, one hundred times. Run by Vals AI.

As of 19 Sept 2026, Claude Opus 5 leads Vibe Code Bench 1-100 on BenchLeader with 28.5%, ahead of Claude Fable 5.1 at 28.0%, across 18 model configurations with a published result.

Published by
Vals AI
Category
Coding
Index weight
Reference only
Models
18
Data as of
19 Sept 2026

Vals AI (vals.ai).

What the test looks like

Starting from a working app, the agent applies a hundred successive feature requests; the score is how many it completes without breaking the app.

How it is scored

Requests completed, run by Vals AI.

What to keep in mind

Tests sustained agentic coding rather than one-shot generation.

Top 15
  1. 1Claude Opus 528.5%
  2. 2Claude Fable 5.128.0%
  3. 3GPT-6 Astra (max)27.6%
  4. 4GPT-5.6 Luna (max)22.6%
  5. 5Muse Spark 1.3 (max)20.5%
  6. 6GPT-5.6 Sol (max)20.0%
  7. 7GLM-5.3 (max)20.0%
  8. 8Gemini 3.8 Flash (high)18.8%
  9. 9Kimi K3 (max)18.2%
  10. 10DeepSeek V4 Pro 0813 (max)17.5%
  11. 11DeepSeek V4.1 Flash (high)16.4%
  12. 12GLM-5.3-Flash (max)16.0%
  13. 13GPT-5.6 Terra (max)14.8%
  14. 14Grok 4.6 (high)14.8%
  15. 15Claude Sonnet 513.8%
18 of 18
#
1Claude Opus 5Anthropic28.5%66.72026-09-16
2Claude Fable 5.1Anthropic28.0%64.52026-09-16
3GPT-6 AstramaxOpenAI27.6%71.82026-09-16
4GPT-5.6 LunamaxOpenAI22.6%60.22026-09-16
5Muse Spark 1.3maxMeta20.5%69.32026-09-16
6GPT-5.6 SolmaxOpenAI20.0%68.82026-09-16
7GLM-5.3maxZhipu AIopen ↗20.0%65.82026-09-16
8Gemini 3.8 FlashhighGoogle18.8%64.52026-09-16
9Kimi K3maxMoonshot AIopen ↗18.2%67.22026-09-16
10DeepSeek V4 Pro 0813maxDeepSeekopen ↗17.5%56.02026-09-16
11DeepSeek V4.1 FlashhighDeepSeekopen ↗16.4%–2026-09-16
12GLM-5.3-FlashmaxZhipu AIopen ↗16.0%55.32026-09-16
13GPT-5.6 TerramaxOpenAI14.8%65.02026-09-16
14Grok 4.6highxAI14.8%64.52026-09-16
15Claude Sonnet 5Anthropic13.8%57.22026-09-16
16Qwen3 8maxAlibaba12.8%66.52026-09-16
17MiniMax-M3MiniMaxopen ↗9.2%57.52026-09-16
18Gemini 3.1 ProhighGoogle6.7%60.72026-09-16

Cite as: BenchLeader, “Vibe Code Bench 1-100 leaderboard”, https://www.benchleader.com/benchmarks/vals_vibe_code_bench, data as of 19 Sept 2026.

Vibe Code Bench 1-100: questions

What does Vibe Code Bench 1-100 measure?
Starting from a working app, the agent applies a hundred successive feature requests; the score is how many it completes without breaking the app. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads Vibe Code Bench 1-100?
Claude Opus 5 leads Vibe Code Bench 1-100 with 28.5% as of 19 Sept 2026, ahead of Claude Fable 5.1 at 28.0%.
How many models have Vibe Code Bench 1-100 results?
18 model configurations have a Vibe Code Bench 1-100 result on BenchLeader, all taken from Vals AI.
Who runs Vibe Code Bench 1-100 and how often is it updated?
Vibe Code Bench 1-100 is published by Vals AI. BenchLeader re-reads the published results every morning and records the date each result was published.
Does Vibe Code Bench 1-100 count toward the BenchLeader Index?
No. Vibe Code Bench 1-100 is shown for reference but left out of the composite index.