AI Radar

Benchmarks

Top models on public leaderboards whose licence allows republishing, as published by each source. Scores are not comparable across boards. Δ7d is the change since the same board a week earlier.

9 boards Collected Nov 10, 15:00 UTC Compare models benchmarks.json

Epoch Capabilities Index

Composite capability score fitted across many benchmarks.

Metric
ECI (ECI), higher is better
Source
Epoch AI, CC-BY-4.0
Updated
Nov 9

Epoch AI

#ModelProviderECIBarΔ7d
1Claude Fable 5.2Anthropic174.2–
2GPT-6 AstraOpenAI169.2–
3GPT-6.1 SolOpenAI168.9–
4Claude Opus 5.5Anthropic168.7–
5GPT-6.1 TerraOpenAI168.1–
6Claude Fable 5.1Anthropic167.0–
7Gemini 4 ArgonGoogle166.5–
8Claude Sonnet 5.5Anthropic166.3–
9Claude Opus 5Anthropic164.5–
10Claude Fable 5Anthropic164.2–
Ranks 11–25
#ModelProviderECIBarΔ7d
11GPT-6 SolOpenAI164–
12GPT-5.6 SolOpenAI163.0–
13GPT-5.5 ProOpenAI162.8–
14Kimi K3.1 ProMoonshot AI162.0–
15Muse WonderMeta162.0–
16GPT-5.6 TerraOpenAI160.9–
17Claude Haiku 5.5Anthropic160.2–
18GPT-5.4 ProOpenAI160.2–
19GPT-5.5OpenAI160.1–
20Claude Opus 4.8Anthropic159.9–
21DeepSeek V4.1 Flash 1017DeepSeek159.3–
22Gemini 3.8 FlashGoogle158.7–
23Kimi K3Moonshot AI158.5–
24Gemini 3.7 FlashGoogle158.4–
25Muse Spark 1.3Meta158.0–

GPQA Diamond

Graduate-level multiple-choice science questions (Epoch's own runs).

Metric
accuracy (%), higher is better
Source
Epoch AI, CC-BY-4.0
Updated
Nov 6

Epoch AI

#ModelProvider%BarΔ7d
1Gemini 4 ArgonGoogle96.4–
2GPT-6 AstraOpenAI95.8–
3Claude Sonnet 5.5Anthropic95.6–
4Gemini 3.8 FlashGoogle95.4–
5GPT-6.1 SolOpenAI95.4–
6Gemini 3.7 FlashGoogle94.8–
7GPT-5.4 ProOpenAI94.6–
8Gemini 3.1 Pro PreviewGoogle94.4–
9GPT-6 SolOpenAI94.3–
10GPT-6.1 TerraOpenAI94.3–
Ranks 11–25
#ModelProvider%BarΔ7d
11Kimi K3.1 ProMoonshot AI94.2–
12Gemini 3.6 FlashGoogle94.1–
13GPT-5.5OpenAI94.0–
14Grok 4.6xAI94.0–
15GPT-5.5 ProOpenAI93.9–
16Claude Fable 5.2Anthropic93.9–
17Claude Opus 5Anthropic93.9–
18GPT-5.6 SolOpenAI93.5–
19Grok 4.5xAI93.4–
20GPT-5.6 TerraOpenAI93.3–
21GPT-5.4OpenAI93.3–
22DeepSeek V4.1 Flash 1017DeepSeek93.2–
23Doubao Seed EvolvingBytedance93.1–
24Kimi K3Moonshot AI93.1–
25Gemini 3.5 FlashGoogle92.8–

FrontierMath (Tiers 1-3)

Unpublished research-level math problems (Epoch's own runs).

Metric
accuracy (%), higher is better
Source
Epoch AI, CC-BY-4.0
Updated
Nov 6

Epoch AI

#ModelProvider%BarΔ7d
1Claude Fable 5.2Anthropic95.1–
2GPT-6 AstraOpenAI93.7–
3GPT-6.1 SolOpenAI93.7–
4Claude Opus 5.5Anthropic91.2–
5GPT-6.1 TerraOpenAI90.9–
6Claude Fable 5.1Anthropic90.2–
7GPT-6 SolOpenAI89.8–
8GPT-5.6 SolOpenAI89.1–
9Claude Sonnet 5.5Anthropic88.8–
10GPT-5.5 ProOpenAI87.7–
Ranks 11–25
#ModelProvider%BarΔ7d
11Claude Fable 5Anthropic87.0–
12Gemini 4 ArgonGoogle86.0–
13GPT-5.6 TerraOpenAI86.0–
14Claude Opus 5Anthropic85.6–
15GPT-5.5OpenAI85.3–
16GPT-5.4 ProOpenAI82.5–
17GPT-5.6 LunaOpenAI82.1–
18Muse WonderMeta82.1–
19Claude Opus 4.8Anthropic80.0–
20GPT-6 LunaOpenAI78.9–
21GPT-5.4OpenAI78.6–
22Kimi K3.1 ProMoonshot AI78.2–
23Claude Haiku 5.5Anthropic75.1–
24Qwen3.8 MaxQwen74.7–
25Muse Spark 1.3Meta74.4–

SWE-bench Verified

Real GitHub issues the model must fix so the repo's tests pass (Epoch's own runs).

Metric
resolved (%), higher is better
Source
Epoch AI, CC-BY-4.0
Updated
Jun 25

Epoch AI

#ModelProvider%BarΔ7d
1Claude Opus 4.7Anthropic83.5–
2GPT-5.5OpenAI80.6–
3Gemini 3.5 FlashGoogle79.3–
4Claude Opus 4.6Anthropic78.7–
5GLM-5.2Z.ai78.7–
6DeepSeek v4 ProDeepSeek77.6–
7Qwen3.7 MaxQwen77.3–
8GPT-5.4OpenAI76.9–
9Claude Opus 4.5Anthropic76.7–
10Kimi K2.6Moonshot AI76.7–
Ranks 11–25
#ModelProvider%BarΔ7d
11Qwen 3.6 MaxQwen76.7–
12Gemini 3.1 Pro PreviewGoogle75.6–
13Gemini 3 Flash PreviewGoogle75.4–
14Claude Sonnet 4.6Anthropic75.2–
15GPT-5.3 CodexOpenAI74.8–
16GLM-5.1Z.ai74.2–
17GPT-5.2OpenAI73.8–
18Kimi K2.5Moonshot AI73.8–
19GPT-5OpenAI73.6–
20claude-opus-4-1-20250805Anthropic73.3–
21Gemini 3 Pro PreviewGoogle72.9–
22GLM-5Z.ai72.1–
23Claude Sonnet 4.5Anthropic71.3–
24Claude Opus 4Anthropic70.7–
25GPT-5.1OpenAI68.0–

Terminal-Bench

Agent completes real tasks in a terminal; best agent harness per model.

Metric
accuracy (%), higher is better
Source
Terminal-Bench via Epoch AI, CC-BY-4.0
Updated
Apr 23

Terminal-Bench, via Epoch AI

#ModelProvider%BarΔ7d
1GPT-5.5OpenAI84.7–
2GPT-5.4OpenAI81.8–
3Claude Opus 4.7Anthropic80.2–
4Gemini 3.1 Pro PreviewGoogle80.2–
5Claude Opus 4.6Anthropic79.8–
6GPT-5.3 CodexOpenAI78.4–
7Gemini 3 Pro PreviewGoogle69.4–
8GPT-5.2 CodexOpenAI66.5–
9GPT-5.2OpenAI64.9–
10Gemini 3 Flash PreviewGoogle64.3–
Ranks 11–25
#ModelProvider%BarΔ7d
11Claude Opus 4.5Anthropic63.1–
12GPT-5.1 Codex MiniOpenAI61.6–
13GPT-5.1-Codex-MaxOpenAI60.4–
14GPT-5.1 CodexOpenAI57.8–
15Grok 4.20 ReasoningxAI57.3–
16Claude Sonnet 4.6Anthropic53.4–
17GLM-5Z.ai52.4–
18GPT-5OpenAI49.6–
19GPT-5.1OpenAI47.6–
20Claude Sonnet 4.5Anthropic46.5–
21MiniMax M2.7MiniMax45.1–
22GPT-5-codexOpenAI44.3–
23Kimi K2.5Moonshot AI43.2–
24MiniMax M2.7 HighspeedMiniMax42.9–
25minimax-m2.5MiniMax42.7–

Humanity's Last Exam

Very hard expert-written questions across many fields.

Metric
accuracy (%), higher is better
Updated
Oct 7

Humanity's Last Exam (CAIS & Scale AI), via Epoch AI

#ModelProvider%BarΔ7d
1GPT-6 AstraOpenAI54.8–
2Claude Fable 5.1Anthropic46.5–
3Gemini 3.1 Pro PreviewGoogle46.4–
4Claude Sonnet 5.5Anthropic46.1–
5Gemini 3.8 FlashGoogle44.5–
6GPT-5.4 ProOpenAI44.3–
7Muse SparkMeta40.6–
8Gemini 3 Pro PreviewGoogle37.5–
9GPT-5.4OpenAI36.2–
10Claude Opus 4.7Anthropic36.2–
Ranks 11–25
#ModelProvider%BarΔ7d
11Claude Opus 4.6Anthropic34.4–
12GPT-5 ProOpenAI31.6–
13Claude Haiku 5.5Anthropic30.1–
14GPT-5.2OpenAI27.8–
15GPT-5OpenAI25.3–
16Claude Opus 4.5Anthropic25.2–
17Kimi K2.5Moonshot AI24.4–
18GPT-5.1OpenAI23.7–
19Gemini 2.5 Pro PreviewGoogle21.6–
20o3OpenAI20.3–
21GPT-5 miniOpenAI19.4–
22Gemini 2.5 Pro ExpGoogle18.2–
23o4-miniOpenAI18.1–
24Claude Sonnet 4.5Anthropic13.7–
25Gemini 2.5 Flash PreviewGoogle12.1–

ARC-AGI-2

Novel visual pattern puzzles that test fluid reasoning.

Metric
score (%), higher is better
Source
ARC Prize Foundation via Epoch AI, CC-BY-4.0
Updated
Nov 9

ARC Prize Foundation, via Epoch AI

#ModelProvider%BarΔ7d
1GPT-6 AstraOpenAI95.0–
2GPT-6.1 SolOpenAI94.2–
3Claude Fable 5.2Anthropic93.8–
4Claude Opus 5.5Anthropic93.3–
5GPT-5.6 SolOpenAI92.5–
6Gemini 4 ArgonGoogle91.2–
7Claude Opus 5Anthropic90.4–
8Claude Fable 5.1Anthropic90.0–
9GPT-6 SolOpenAI89.6–
10Claude Fable 5Anthropic89.2–
Ranks 11–25
#ModelProvider%BarΔ7d
11Gemini 3.8 FlashGoogle89.2–
12GPT-5.5OpenAI85.0–
13Gemini 3.7 FlashGoogle84.6–
14Gemini 3 Deep ThinkGoogle84.6–
15GPT-5.5 ProOpenAI84.6–
16GPT-5.6 TerraOpenAI83.9–
17Claude Sonnet 5.5Anthropic83.9–
18GPT-5.4 ProOpenAI83.3–
19Inkling 2 SmallThinking Machines81.7–
20DeepSeek V4.1 Flash 1017DeepSeek78.1–
21Gemini 3.1 Pro PreviewGoogle77.1–
22Claude Opus 4.7Anthropic75.8–
23GPT-5.4OpenAI74.0–
24DeepSeek V4.1 FlashDeepSeek72.9–
25Claude Opus 4.8Anthropic72.1–

LMArena Text

Human preference rating from blind side-by-side chats (style-controlled).

Metric
rating (rating), higher is better
Source
LMArena, CC-BY-4.0
Updated
Nov 9

LMArena

#ModelProviderratingBarΔ7d
1Gemini 4 Carbon (High)Google1526–
2Gemini 4 Argon (High)Google1522.4–
3Muse Wonder (Max)Meta1511–
4Claude Opus 5.5 (High)Anthropic1506.1–
5Claude Fable 5 (High)Anthropic1503.9–
6Claude Opus 4.6 (High)Anthropic1503–
7Claude Fable 5.1 (Max)Anthropic1502.7–
8Claude Opus 4.7 (High)Anthropic1501.1–
9Gemini 3.8 Flash (High)Google1497.5–
10Muse Spark 1.3 (Max)Meta1496.3–
Ranks 11–25
#ModelProviderratingBarΔ7d
11Muse Spark 1.2 (xHigh)Meta1493.5–
12Muse Spark 1.1Meta1490.7–
13Claude Opus 5 (High)Anthropic1489.8–
14Gemini 3.7 Flash (High)Google1489.4–
15Muse SparkMeta1489.1–
16Kimi K3 (Max)Moonshot AI1487.1–
17Gemini 3.1 Pro PreviewGoogle1486.7–
18Gemini 3 ProGoogle1485.6–
19GPT-5.6 Sol (xHigh)OpenAI1484.9–
20Qwen3.8-MaxQwen1484–
21Kimi K3.1 Pro (Max)Moonshot AI1483–
22Gemini 3.6 Flash (High)Google1482.5–
23GLM-5.3 (Max)Z.ai1482.5–
24GPT-6.1 Sol (Max)OpenAI1482.1–
25Claude Opus 4.8 (High)Anthropic1481.8–

LMArena WebDev

Human preference rating for building web apps from a prompt.

Metric
rating (rating), higher is better
Source
LMArena, CC-BY-4.0
Updated
Nov 9

LMArena

#ModelProviderratingBarΔ7d
1Claude Opus 5.5 (Max)Anthropic1817.7–
2Claude Sonnet 5.5 (xHigh)Anthropic1774.3–
3Gemini 4 Carbon (High)Google1768–
4GPT-6 Astra (Max)OpenAI1766.8–
5Claude Fable 5.1 (Max)Anthropic1744.6–
6GPT-6.1 Sol (Max)OpenAI1742.8–
7Muse Wonder Full (Max)Meta1723.9–
8Claude Opus 5 (Max)Anthropic1690.8–
9Kimi K3.1 Pro (Max)Moonshot AI1687.8–
10Gemini 4 Argon (High)Google1682.6–
Ranks 11–25
#ModelProviderratingBarΔ7d
11GPT-6.1 Terra (Max)OpenAI1678.7–
12Qwen3.8-Max (0902)Qwen1676.7–
13GPT-6 Sol (Max)OpenAI1673.9–
14DeepSeek-V4.1-Flash-20261017 (Max)DeepSeek1664.2–
15Kimi K3 (Max)Moonshot AI1657.1–
16Muse Spark 1.3 (Max)Meta1654.3–
17Muse Wonder (Max)Meta1650.3–
18Inkling 2 SmallThinky1649.5–
19Grok 4.7 (xHigh)xAI1639.4–
20MiMo-V2.6-FlashXiaomi1636.5–
21Hy4 previewTencent1631.5–
22Claude Fable 5 (High)Anthropic1625.2–
23Qwen3.8-Flash-NextQwen1624.5–
24MiMo-V2.6-ProXiaomi1623.9–
25GLM-5.3 (Max)Z.ai1622.2–

Compare models

Each model's score on every board, with list price (USD per 1M tokens) and context length. The best score in each column is underlined; a dash means the board does not list the model. Showing 47 of 88 models: those listed on the most boards. All are in benchmarks.json.

Model Provider Epoch Capabilities IndexECI GPQA Diamond% FrontierMath (Tiers 1-3)% SWE-bench Verified% Terminal-Bench% Humanity's Last Exam% ARC-AGI-2% LMArena Textrating LMArena WebDevrating $ in $ out Context
Claude Fable 5.1Anthropic167.0–90.2––46.590.01502.71744.6–––
Claude Opus 5Anthropic164.593.985.6–––90.41489.81690.8–––
Claude Sonnet 5.5Anthropic166.395.688.8––46.183.9–1774.3–––
Gemini 3.1 Pro PreviewGoogle–94.4–75.680.246.477.11486.7––––
Gemini 4 ArgonGoogle166.596.486.0–––91.21522.41682.6–––
GPT-5.4OpenAI–93.378.676.981.836.274.0–––––
GPT-5.5OpenAI160.194.085.380.684.7–85.0–––––
GPT-6.1 SolOpenAI168.995.493.7–––94.21482.11742.8–––
GPT-6 AstraOpenAI169.295.893.7––54.895.0–1766.8–––
Claude Fable 5Anthropic164.2–87.0–––89.21503.91625.2–––
Claude Opus 4.7Anthropic–––83.580.236.275.81501.1––––
Claude Opus 5.5Anthropic168.7–91.2–––93.31506.11817.7–––
Gemini 3.8 FlashGoogle158.795.4–––44.589.21497.5––––
Kimi-K3.1-ProMoonshot AI162.094.278.2––––14831687.8–––
GPT-5.4 ProOpenAI160.294.682.5––44.383.3–––––
GPT-5.6 SolOpenAI163.093.589.1–––92.51484.9––––
GPT-6 SolOpenAI16494.389.8–––89.6–1673.9–––
Claude Fable 5.2Anthropic174.293.995.1–––93.8––8.0040.001M
Claude Opus 4.6Anthropic–––78.779.834.4–1503––––
Claude Opus 4.8Anthropic159.9–80.0–––72.11481.8––––
Gemini 3.7 FlashGoogle158.494.8––––84.61489.4––––
Muse Spark 1.3Meta158.0–74.4––––1496.31654.3–––
Muse WonderMeta162.0–82.1––––15111650.3–––
Kimi K3Moonshot AI158.593.1–––––1487.11657.1–––
GPT-5.5 ProOpenAI162.893.987.7–––84.6–––––
GPT-5.6 TerraOpenAI160.993.386.0–––83.9–––––
GPT-6.1 TerraOpenAI168.194.390.9–––––1678.7–––
Claude Haiku 5.5Anthropic160.2–75.1––30.1––––––
Claude Opus 4.5Anthropic–––76.763.125.2––––––
Claude Sonnet 4.5Anthropic–––71.346.513.7––––––
DeepSeek V4.1 Flash 1017DeepSeek159.393.2––––78.1–––––
Gemini 3 Pro PreviewGoogle–––72.969.437.5––––––
Kimi K2.5Moonshot AI–––73.843.224.4––––––
GPT-5OpenAI–––73.649.625.3––––––
GPT-5.1OpenAI–––68.047.623.7––––––
GPT-5.2OpenAI–––73.864.927.8––––––
Qwen3.8 MaxQwen––74.7––––14841676.7–––
Claude Sonnet 4.6Anthropic–––75.253.4–––––––
DeepSeek V4.1 FlashDeepSeek––––––72.9–1664.2–––
Gemini 3.5 FlashGoogle–92.8–79.3––––––––
Gemini 3.6 FlashGoogle–94.1–––––1482.5––––
Gemini 3 Flash PreviewGoogle–––75.464.3–––––––
Gemini 4 Carbon (High)Google–––––––15261768–––
Muse SparkMeta–––––40.6–1489.1––––
GPT-5.3 CodexOpenAI–––74.878.4–––––––
GLM-5Z.ai–––72.152.4–––––––
GLM-5.3 (Max)Z.ai–––––––1482.51622.2–––

Prices and context come from provider listings (via OpenRouter where available) and may lag behind changes.