GPT-5

Standard benchmark by various with evaluation knowledge • Self-reported

Mode thinking (to 128K tokens) with capabilities reasoning and approach to solving problems • Self-reported

Benchmark generation code with by functions on Python • Self-reported

mode thinking with solution and mathematical tasks • Self-reported

GPT-5 - Diamond-thinking without tools This methodology uses mode thinking, which differs from more standard "and " thinking. consists in that, that we we use so "diamond-thinking" (thinking) - method, where model: 1. with on problem, set various approaches and possible 2. to most 3. these on thoroughly evaluating their 4. in full, solution This approach especially efficient for mathematical and tasks, requiring thinking for search solutions. During many cases model achieves exact, correct solutions, which not was would at use more methods thinking. Since we not we tool use, results can directly compare with where tools, in order to better understand their on performance model • Self-reported

GPT-5 with mode thinking - Solution visual tasks level with • Self-reported

mode thinking (to 128 thousand tokens) with reasoning and understanding code on different languages programming • Self-reported

GPT-5 - IC SWE Diamond Freelance Coding Tasks (evaluation on basis ) • Self-reported

GPT-5 standard with included mode thinking (without tools) - mathematics • Self-reported

Mode thinking for in complex on • Self-reported

GPT-5 standard with included mode thinking (only with tool python) - mathematics expert level FrontierMath levels 1-3. • Self-reported

GPT-5 standard with included mode thinking (without tools) - Harvard-MIT Mathematics Tournament. • Self-reported

GPT-5 standard with mode thinking (without tools) - set questions expert level by various subjects • Self-reported

GPT-5 with included mode thinking - Benchmark execution multi-step instructions. • Self-reported

GPT-5 with included mode thinking - Benchmark agentic search and • Self-reported

GPT-5 with included mode thinking - instructions in form • Self-reported

GPT-5 with o3-mini - Benchmark execution instructions with accuracy evaluation • Self-reported

GPT-5 - Evaluation execution instructions through internal API (complexity) • Self-reported

GPT-5 - Benchmark functions () • Self-reported

GPT-5 with mode thinking - Benchmark functions () • Self-reported

GPT-5 with mode thinking - Benchmark functions (field) • Self-reported

GPT-5 with mode thinking - solution visual tasks level with reasoning • Self-reported

GPT-5 with mode thinking - reasoning (256 ) • Self-reported

GPT-5 with mode thinking - Reasoning and scientific • Self-reported

GPT-5 with mode thinking - thinking • Self-reported

OpenAI-MRCR - search at 128 tokens • Self-reported

OpenAI-MRCR 2-needle retrieval at 256 tokens • Self-reported

in (Graphwalks BFS) (<128k) for reasoning with context • Self-reported

on (<128k), for reasoning with context • Self-reported

BrowseComp option with context 128k • Self-reported

BrowseComp option with context 256k • Self-reported

VideoMME (long) with — • Self-reported

mode thinking for on for queries • Self-reported

Mode thinking for on for queries, on • Self-reported

Mode thinking for evaluation actual accuracy. level on • Self-reported

GPT-5 with thinking mode enabled - Multi-turn instruction following benchmark. • Self-reported