ViBench 2.0 - All17 Plan Pass: leaderboard

Long-horizon greenfield software building with sequential features and browser-graded test plans. Browser test-plan pass rate (%). ViBench2 native public OpenHands-SDK harness, medium reasoning, temperature1.0, max300steps and2hours per stage, summarization past200k tokens; provider default output/context limits and recommended open-weight top-p. Independent cohort ran2.0.1.beta, historical-GPT cohort2.0.5 with matching task/plan hashes. Four independent builds per product. Pass@1 requires every browser-graded plan to pass; Plan Pass is the mean fraction of plans passed. Grader/seeding Opus5.5 medium.

Metric: Browser test-plan pass rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked.

Top models

#ModelScoreOverall rank
1GPT-6 Astra (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)98
2GPT-6.1 Sol (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)96.9
3Claude Opus 5.5 (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)93.7
4Claude Sonnet 5.5 (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)79.2
5DeepSeek V4.1 Flash (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)77.4
6GLM-5.3 (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)66
7Kimi K3 (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)56.7
8Grok 4.7 (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)52.9
9Gemini 3.8 Flash (Benchmark Sweep Run: source=ViBench2; OpenHandsSDK; medium; temperature1.0; cohort=independent17)38.1

Interactive version: theaggregate.ai/benchmark?slug=vibench-2-0-all17-plan-pass · How It Works · Data refreshed daily, snapshot 2026-11-06.