AI3

Scores

Review raw scores and compare results within each public benchmark.

Data as of 10/28/2026

9 public scores available

9 public benchmarksUpdated through

LMArena Text

Asks people to compare two anonymous model responses and vote, measuring which models users actually prefer.

Latest score: · 22 brands

Scroll sideways to see all 22 brands.

View chart data
DateBrandModel / evaluation configRaw score
ClaudeClaude Opus 5.5 · High1509
GeminiGemini 3.8 Flash · High1496
MuseMuse Spark 1.2 · XHigh1492
KimiKimi K3 · Max1488
GPTGPT-5.6 Sol · XHigh1485
QwenQwen 3.8 Max1482
GLMGLM-5.3 · Max1481
GrokGrok 4.20 Beta 11475
ERNIEERNIE 5.11467
MiMoMiMo-V2.5-Pro1467
DeepSeekDeepSeek V4 Pro1458
MiniMaxMiniMax M31439
MistralMistral Large 31414
StepStep 3.5 Flash1393
o-serieso1-preview1389
HunyuanHunyuan T11387
LlamaLlama 3.3 70B Instruct1318
AtheneAthene 70B1307
RekaReka Core1288
OLMoOLMo 2 32B Instruct1252
OpenChatOpenChat 3.51183
PhiPhi-3 Mini 4K Instruct1128

LiveBench Agentic Coding

Uses coding tasks that require autonomous planning and execution to measure a model's ability to solve code problems independently.

Latest score: · 10 brands

Scroll sideways to see all 10 brands.

View chart data
DateBrandModel / evaluation configRaw score
QwenQwen 3.8 Max64.65%
KimiKimi K362.17%
GLMGLM-5.360.91%
MuseMuse Spark 1.1 · XHigh58.54%
GeminiGemini 3.7 Flash · High58.28%
GrokGrok 4.657.02%
GPTGPT-5.6 Sol · Max56.21%
DeepSeekDeepSeek V4 Pro54.95%
ClaudeClaude Sonnet 4.5 · Thinking 64K53.33%
MiniMaxMiniMax M340.66%

τ³ Banking Knowledge

Uses knowledge retrieval, long conversations and tool calls in banking support to measure whether agents can complete realistic requests.

Latest score: · 8 brands

Scroll sideways to see all 8 brands.

View chart data
DateBrandModel / evaluation configRaw score
QwenQwen 3.8 Max · XHigh55.15%
ClaudeClaude Opus 5 · Max48.71%
GrokGrok 4.5 · High47.94%
GPTGPT-5.6 Sol · XHigh46.91%
MuseMuse Spark 1.1 · XHigh40.46%
GLMGLM-5.2 · XHigh37.11%
KimiKimi K3 · Max37.11%
GeminiGemini 3.1 Pro Preview · High26.03%

GPQA-Diamond

Uses difficult biology, physics and chemistry questions written by PhD-level experts to measure advanced scientific knowledge and reasoning.

Latest score: · 13 brands

Scroll sideways to see all 13 brands.

View chart data
DateBrandModel / evaluation configRaw score
GPTGPT-6 Astra · Max95.77%
GeminiGemini 3.8 Flash · High95.39%
DeepSeekDeepSeek V4.1 Flash 1017 · Max93.18%
GrokGrok 4.6 · XHigh93.18%
KimiKimi K3 · High91.92%
GLMGLM-5.3 · Max90.91%
MiniMaxMiniMax M390.91%
QwenQwen 3.7 Max90.91%
MuseMuse Spark89.8%
ClaudeClaude Haiku 5.5 · Max89.65%
InklingInkling Small · XHigh88.51%
NemotronNemotron 3 Ultra85.35%
o-serieso3 · Medium80.81%

OTIS Mock AIME 2024-2025

Uses challenging mathematics-competition problems to measure mathematical reasoning and problem solving.

Latest score: · 13 brands

Scroll sideways to see all 13 brands.

View chart data
DateBrandModel / evaluation configRaw score
GPTGPT-6 Astra · Max100%
DeepSeekDeepSeek V4.1 Flash 1017 · Max99.44%
GrokGrok 4.6 · XHigh99.17%
ClaudeClaude Haiku 5.5 · XHigh98.89%
GeminiGemini 3.8 Flash · High98.89%
KimiKimi K2.7 Code95.56%
QwenQwen 3.7 Max95.56%
GLMGLM-5.3 · Max91.11%
InklingInkling Small · XHigh90%
MuseMuse Spark88.9%
NemotronNemotron 3 Ultra86.67%
o-serieso3 · Medium84.44%
MiniMaxMiniMax M371.11%

AIME 2025 · OpenCompass

Uses challenging mathematics-competition problems to measure mathematical reasoning and problem solving.

Latest score: · 8 brands

Scroll sideways to see all 8 brands.

View chart data
DateBrandModel / evaluation configRaw score
GLMGLM-595.8%
StepStep 3.5 Flash95.7%
GPTGPT-5 · High94.1%
GeminiGemini 3 Pro Preview93.4%
KimiKimi K2.591.9%
o-serieso4-mini · High91%
DeepSeekDeepSeek R189%
ClaudeClaude Sonnet 4 · Thinking68.7%

SWE-bench Verified

Uses GitHub issues and tests from real open-source projects to measure whether models can understand a repository and implement correct fixes.

Latest score: · 8 brands

Scroll sideways to see all 8 brands.

View chart data
DateBrandModel / evaluation configRaw score
ClaudeClaude Opus 4.7 · Max83.47%
GPTGPT-5.5 Pre-release · XHigh80.58%
GeminiGemini 3.5 Flash · High79.34%
GLMGLM-5.2 · Max78.7%
DeepSeekDeepSeek V4 Pro · Max77.64%
QwenQwen 3.7 Max77.27%
KimiKimi K2.676.65%
o-serieso3 · Medium62.32%

LiveCodeBench V6 · OpenCompass

Uses newly collected programming-contest problems to measure test-passing code generation while reducing training-data contamination.

Latest score: · 8 brands

Scroll sideways to see all 8 brands.

View chart data
DateBrandModel / evaluation configRaw score
GLMGLM-586.2%
StepStep 3.5 Flash83.9%
GPTGPT-5 · High83.2%
GeminiGemini 3 Pro Preview82.9%
KimiKimi K2.580.6%
o-serieso372.7%
DeepSeekDeepSeek R161%
ClaudeClaude Sonnet 4 · Thinking47.5%

Terminal-Bench 2.1

Uses software, system and data tasks in real terminal environments to measure agents' ability to operate computers independently.

Latest score: · 6 brands

Scroll sideways to see all 6 brands.

View chart data
DateBrandModel / evaluation configRaw score
GrokGrok 4.5 · Cursor CLI · High79.33%
ClaudeClaude Opus 4.8 · Claude Code · High78.88%
GPTGPT-5.6 Terra · Codex · Max78.43%
MuseMuse Spark 1.1 · mini-SWE-agent · XHigh76.18%
GeminiGemini 3.1 Pro Preview · Gemini CLI · High65.84%
GLMGLM-5.1 · Claude Code · Max58.65%