Atlas

Models

← All models

GPT-4

OpenAI · 2023-03-14

GPT-4 is OpenAI's older high-intelligence GPT model, released through the API as the text-only gpt-4 model ID with an 8,192-token context window. The broader GPT-4 system was developed as a large multimodal model and scored around the top 10% of test takers on a simulated bar exam; OpenAI did not disclose its model size or detailed architecture.

Benchmark scores

BenchmarkScore
Agentic Tool Use (Chat)37.9
Agentic Tool Use (Enterprise)51.4
AidanBench909
AlpacaEval 2.0 (length-controlled win rate)38.1
AlpacaEval 2.0 (raw win rate)23.6
ARC AI296.3
Arena Text — Business, Management & Financial Ops (No Style Control)1145
Arena Text — Chinese (No Style Control)1183
Arena Text — Creative Writing (No Style Control)1191
Arena Text — English (No Style Control)1220
Arena Text — Entertainment, Sports & Media (No Style Control)1185
Arena Text — Exclude Ties (No Style Control)1086
Arena Text — Exclude Ties (Style Controlled)1201
Arena Text — Expert (No Style Control)1148
Arena Text — French (No Style Control)1218
Arena Text — French (Style Controlled)1306
Arena Text — German (No Style Control)1198
Arena Text — Hard Prompts (No Style Control)1200
Arena Text — Hard Prompts English (No Style Control)1211
Arena Text — Instruction Following (No Style Control)1202
Arena Text — Japanese (No Style Control)1137
Arena Text — Korean (No Style Control)1087
Arena Text — Legal & Government (No Style Control)1191
Arena Text — Life, Physical & Social Science (No Style Control)1177
Arena Text — Longer Query (No Style Control)1190
Arena Text — Medicine & Healthcare (No Style Control)1144
Arena Text — Multi-turn (No Style Control)1206
Arena Text — Non-English (No Style Control)1194
Arena Text — Non-English (Style Controlled)1271
Arena Text — Occupational Mathematical (No Style Control)1228
Loading Atlas data…