BenchPress - By Category

Category Heatmap (Top 15)

Composite score (0 to 100) per category for the top 15 overall models. Full leaderboard is on the Generalist page.

Category	gpt-5.3	claude-opus-4.7	claude-sonnet-4.6	claude-opus-4.6	claude-opus-4.5	gpt-5.2	gpt-5.4	claude-sonnet-4.5	gemini-3-flash	gpt-5.1	glm-5	kimi-k2.5	gemini-3-pro	gemini-3.1-pro	gpt-5.5
behavioural	95	95	91	93	91	90	89	91	87	92	89	90	86	85	93
coding	98	96	95	94	94	96	99	92	93	95	91	94	91	91	91
instruction following	97	89	91	86	85	92	90	90	92	85	97	83	92	96	94
learning	99	100	100	100	99	100	99	98	99	100	98	100	99	98	95
meta	85	98	98	93	97	86	86	93	89	94	79	87	93	91	82
reasoning	95	96	95	96	96	94	90	95	96	92	97	96	95	96	95
research	97	95	97	98	98	98	98	95	96	96	95	95	94	98	89
writing	97	95	97	98	97	96	97	98	98	97	98	98	99	90	98

Sycophancy resistance, calibration under social pressure, pushback against confident-but-wrong claims.

Code review, debugging, implementation. Tests pattern recognition, language-specific knowledge, and ability to spot subtle bugs.

Strict format and constraint adherence: exact list lengths, ordered steps, banned words, structural rules.

Explanatory writing on technical topics. Tests how well the model teaches a concept to a target audience.

Multi-step quantitative reasoning, Fermi estimation, logical deduction, statistical analysis.

Open-ended research and synthesis: comparisons, tradeoff analysis, design recommendations.

Production writing (docs, summaries, explanations) with constraints on length, audience, and format.