CoRE (Code Reasoning) - IO: leaderboard

CoRE (2026) tests code reasoning on 60 HumanEval and LiveCodeBench problems, each with several functionally equivalent implementations and probes of intermediate execution states, here with standard input-output prompting. The Reasoning Consistency Score in percent credits a problem only when every implementation is predicted correctly, scaled by the share of intermediate probes answered correctly; higher is better.

Metric: Reasoning Consistency Score (RCS, %) with standard input-output prompting: 60 HumanEval and LiveCodeBench problems with 255 expert-verified functionally equivalent implementations and intermediate-state probes; per problem the score is the strict output consistency (every implementation and test predicted correctly, 0 or 1) times the share of intermediate probes answered correctly, averaged over problems; higher is better. Source: arxiv.org. Saturation forecast: Around November 2027. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1DeepSeek V3.240.49#419
2CoRE Claude-4.5 (checkpoint unspecified)39.45
3DeepSeek R120.6#482
4GPT-515.93#178
5Gemini-2.5 (CoRE checkpoint unspecified)12.11
6O38.9#237
7Claude 3.7 Sonnet6.81#470
8Qwen 3 235B A22B0.27#628

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=core-code-reasoning-io · How It Works · Data refreshed daily, snapshot 2026-11-21.