LLM Benchmarks — Scores, Methodology & Live Rankings
What each benchmark measures, how models score, and the current rankings
Terminal-Bench 2.1 (Best Reported Harness)
Agentic Coding
Terminal-Bench 2.1 (Best Reported Harness) - best score across all evaluation harnesses (Terminus-2, Claude Code, etc.) for Terminal-Bench 2.1.
CodeForces
STEM & Reasoning
CodeForces - competitive programming rating evaluated on custom query set.
FullStackBench en
Agentic Coding
FullStackBench en - English full-stack code generation benchmark.
FullStackBench zh
Agentic Coding
FullStackBench zh - Chinese full-stack code generation benchmark.
SUNRGBD
Spatial Intelligence
SUNRGBD - 3D spatial understanding benchmark using RGB-D indoor scenes.
OCRBench
Document Understanding
OCRBench - comprehensive OCR capability evaluation benchmark.
HLE with search
STEM & Reasoning
HLE with search - Humanity's Last Exam with web search access variant.
BigBench Extra Hard
STEM & Reasoning
BigBench Extra Hard - extended and more challenging version of BigBench.
CoVoST
Multilingual
CoVoST - Conversational Voice Translation benchmark for speech-to-text translation across multiple languages.
FLEURS
Multilingual
FLEURS - Few-shot Learning Evaluation of Universal Representations of Speech. Lower is better (error rate).
MRCR v2 128K (8-needle)
Long Context
MRCR v2 128K (8-needle) - Multi-Ring Curve Retrieval benchmark v2, 128K context, 8-needle setting.
DeepSWE
Agentic Coding
DeepSWE - agentic coding benchmark from datacurve/deep-swe dataset evaluating software engineering agent capabilities.
DSBench-FullStack
Agentic Coding
DSBench-FullStack - internal full-stack development test set for evaluating coding agents on full-stack development tasks.
ProgramBench
Agentic Coding
ProgramBench - agentic coding benchmark evaluated with Claude-Code 2.1.156, 200 instances, temp=1.0, top_p=1.0, max_tokens=64000, max_turns=2000, sample_time...
BrowseComp-zh
General Agents
BrowseComp-zh - Chinese web browsing and comprehension agentic benchmark.
Cybergym
General Agents
Cybergym - agentic cybersecurity benchmark evaluating autonomous hacking and security task capabilities.
Aider
Agentic Coding
Aider - coding benchmark based on the Aider coding assistant tool
Frontier-Bench v0.1
Agentic Coding
Frontier-Bench v0.1 - frontier coding benchmark evaluating advanced agentic coding capabilities.
Kimi Code Bench V2
Agentic Coding
In-house benchmark by Moonshot AI designed to evaluate coding agents on realistic tasks. It has diverse software engineering tasks across 10+ mainstream prog...
Kimi Claw 24/7 Bench
Agentic
In-house benchmark by Moonshot AI for evaluating long-horizon agentic performance in persistent, multi-day coworking tasks. It spans 17 professional scenario...
MCPMark-Verified
Agentic
A human-verified edition of MCPMark, a benchmark for evaluating MCP tool use across five real server environments — Notion, GitHub, Filesystem, Postgres, and...
LHTB Solved
Agentic Coding
Long-Horizon-Terminal-Bench (LHTB) Solved metric from IntelligenceLab. Evaluates long-horizon terminal-based agentic tasks.
BrowseComp Agent Swarm
General Agents
BrowseComp evaluated with agent swarm mode - multiple coordinated sub-agents performing web browsing tasks.
HealthBench Consensus
General Capabilities
HealthBench Consensus - a subset of HealthBench validated by the consensus of multiple physicians.
AIME 2024 (with tools)
STEM & Reasoning
AIME 2024 (with tools) - American Invitational Mathematics Examination 2024 with tool use variant.
HealthBench Hard
General Capabilities
HealthBench Hard - a challenging subset of HealthBench conversations testing realistic health conversations with individuals and health professionals.
MMLU
Knowledge
MMLU (Massive Multitask Language Understanding) - a benchmark of multitask accuracy across 57 tasks.
Tau-Bench Airline
General Agents
τ-Bench Airline - function calling ability benchmark in an airline domain. Measures tool use with developer-defined function schemas.
Tau-Bench Retail
General Agents
τ-Bench Retail - function calling ability benchmark in a retail domain. Measures tool use with developer-defined function schemas.
AIME 2025
STEM & Reasoning
AIME 2025 (American Invitational Mathematics Examination) - math competition problems evaluated pass@1 without tools.
DSBench-Hard
Agentic Coding
DSBench-Hard - internal test set of difficult coding-agent problems for evaluating agentic coding on challenging software engineering tasks.
AIME 2024
STEM & Reasoning
AIME 2024 (American Invitational Mathematics Examination) - math competition problems evaluated pass@1 without tools.
Program Bench
Agentic Coding
Evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and its documentation. It spans 200 tasks, from ...
GPQA Diamond (with tools)
STEM & Reasoning
GPQA Diamond (with tools) - Google-Proof Q&A benchmark for graduate-level scientific reasoning with tool use variant.