The top score is 100. Scores a few points apart are close, because tests often disagree. A model with few tests yet (scored in fewer than 3 of the 6 jobs, or on too few tests) is listed without a place: its score may still move. Prices are official list prices in US dollars per million tokens.
Price$2 / $10in / out, per million tokens, until November 30. Then $4 / $20.
Each job on its own · scored in 1 of 6
Chat & writing100.0
Its tests
Gemini 4 Argon took 7 of our tests, from 2 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
76%
1 of 32
27%
Arena AgentArena (LMArena)
+13.59 points
5 of 34
20%
Arena WebDevArena (LMArena)
1678 rating
8 of 37
17%
Arena TextArena (LMArena)
1525 rating
1 of 34
14%
Furniture AssemblyEpoch AI
72.5%
5 of 28
13%
GPQA DiamondEpoch AI
96.4%
1 of 32
4%
Chess PuzzlesEpoch AI
59%
6 of 35
4%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it. Few tests yet: it has a job score of its own in only 1 of the 6 jobs, so the score may still move. It is listed, but gets no place.
AnthropicClosed model
Claude Opus 5.5
Place#1of 33 models with enough tests
Silicon score97.495.6 to 99.2 with other tests
Price$4 / $20in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing88.2
Coding100.0
Math & reasoning96.3
AI agents100.0
Web research90.7
Understanding images100.0
Its tests
Claude Opus 5.5 took 25 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
72.2%
5 of 32
10%
Arena SearchArena (LMArena)
1213 rating
12 of 17
9%
Arena VisionArena (LMArena)
1293 rating
7 of 29
9%
Arena AgentArena (LMArena)
+17.01 points
1 of 34
7%
Arena WebDevArena (LMArena)
1811 rating
1 of 37
6%
MirrorCodeEpoch AI
77.4%
1 of 8
6%
Arena TextArena (LMArena)
1508 rating
2 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
86.52%
4 of 13
4%
LiveBench ReasoningLiveBench
92.15%
3 of 34
4%
EBR-benchEpoch AI
71.43%
2 of 17
4%
Furniture AssemblyEpoch AI
83.33%
1 of 28
4%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
48.37%
4 of 13
4%
LiveBench LanguageLiveBench
86.27%
9 of 34
3%
LiveBench Instruction FollowingLiveBench
67.05%
23 of 34
3%
LiveBench CodingLiveBench
89.25%
2 of 34
3%
LiveBench Agentic CodingLiveBench
71.72%
2 of 34
3%
ArXivMathMathArena (ETH Zürich)
83.65%
4 of 11
3%
BrokenArXivMathArena (ETH Zürich)
90.79%
2 of 11
3%
FrontierMath Tiers 1–3Epoch AI
91.23%
3 of 34
2%
Mystery Game PuzzlesEpoch AI
71%
3 of 32
2%
LiveBench MathematicsLiveBench
97.08%
1 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
68.46%
12 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
57.87%
8 of 14
1%
GPQA DiamondEpoch AI
90.59%
24 of 32
1%
Chess PuzzlesEpoch AI
67%
3 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-6.1 Sol
Place#2of 33 models with enough tests
Silicon score96.594.6 to 98.4 with other tests
Price$2 / $10in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing89.4
Coding88.4
Math & reasoning99.8
AI agents94.9
Web research95.6
Understanding images100.0
Its tests
GPT-6.1 Sol took 24 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
73.9%
3 of 32
10%
Arena SearchArena (LMArena)
1233 rating
8 of 17
9%
Arena VisionArena (LMArena)
1294 rating
6 of 29
9%
Arena AgentArena (LMArena)
+15.05 points
2 of 34
8%
Arena WebDevArena (LMArena)
1760 rating
4 of 37
6%
Arena TextArena (LMArena)
1484 rating
14 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
88.16%
3 of 13
5%
Furniture AssemblyEpoch AI
80%
2 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
48.46%
3 of 13
5%
LiveBench ReasoningLiveBench
92.63%
2 of 34
4%
EBR-benchEpoch AI
54.29%
4 of 17
4%
LiveBench LanguageLiveBench
90.13%
2 of 34
3%
LiveBench Instruction FollowingLiveBench
74.15%
9 of 34
3%
LiveBench CodingLiveBench
80.74%
14 of 34
3%
LiveBench Agentic CodingLiveBench
56.77%
16 of 34
3%
ArXivMathMathArena (ETH Zürich)
95.31%
1 of 11
3%
BrokenArXivMathArena (ETH Zürich)
92.1%
1 of 11
3%
FrontierMath Tiers 1–3Epoch AI
93.68%
1 of 34
2%
Mystery Game PuzzlesEpoch AI
80%
2 of 32
2%
LiveBench MathematicsLiveBench
96.83%
3 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
82.42%
2 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
82.11%
2 of 14
1%
GPQA DiamondEpoch AI
95.39%
4 of 32
1%
Chess PuzzlesEpoch AI
68%
2 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Fable 5.1
Place#3of 33 models with enough tests
Silicon score96.594.6 to 98.4 with other tests
Price$10 / $50in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing90.4
Coding94.2
Math & reasoning91.8
AI agents94.4
Web research97.1
Understanding images99.4
Its tests
Claude Fable 5.1 took 21 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
70.8%
6 of 32
11%
Arena SearchArena (LMArena)
1246 rating
5 of 17
10%
Arena AgentArena (LMArena)
+14.32 points
3 of 34
8%
Arena WebDevArena (LMArena)
1743 rating
5 of 37
7%
MirrorCodeEpoch AI
73.3%
2 of 8
7%
Arena TextArena (LMArena)
1501 rating
6 of 34
6%
Arena VisionArena (LMArena)
1286 rating
12 of 29
5%
Arena DocumentArena (LMArena)
1513 rating
2 of 17
5%
Furniture AssemblyEpoch AI
70%
6 of 28
5%
LiveBench ReasoningLiveBench
91.69%
4 of 34
4%
EBR-benchEpoch AI
57.14%
3 of 17
4%
LiveBench LanguageLiveBench
89.5%
3 of 34
3%
LiveBench Instruction FollowingLiveBench
72.99%
11 of 34
3%
LiveBench CodingLiveBench
86.38%
3 of 34
3%
LiveBench Agentic CodingLiveBench
66.06%
3 of 34
3%
FrontierMath Tiers 1–3Epoch AI
90.18%
4 of 34
3%
Mystery Game PuzzlesEpoch AI
58%
6 of 32
3%
ArXivMathMathArena (ETH Zürich)
56.25%
9 of 11
3%
BrokenArXivMathArena (ETH Zürich)
80.85%
5 of 11
3%
LiveBench MathematicsLiveBench
97.01%
2 of 34
2%
Chess PuzzlesEpoch AI
48%
14 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-6 Astra
Place#4of 33 models with enough tests
Silicon score95.994.1 to 97.7 with other tests
Price$10 / $50in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing88.8
Coding88.0
Math & reasoning100.0
AI agents96.7
Web research97.0
Understanding images95.2
Its tests
GPT-6 Astra took 26 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
75.6%
2 of 32
10%
Arena SearchArena (LMArena)
1234 rating
7 of 17
9%
Arena AgentArena (LMArena)
+13.69 points
4 of 34
7%
Arena WebDevArena (LMArena)
1780 rating
2 of 37
6%
MirrorCodeEpoch AI
46.7%
4 of 8
6%
Arena TextArena (LMArena)
1477 rating
20 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
89.92%
1 of 13
4%
LiveBench ReasoningLiveBench
92.65%
1 of 34
4%
EBR-benchEpoch AI
76.19%
1 of 17
4%
Arena VisionArena (LMArena)
1286 rating
13 of 29
4%
Arena DocumentArena (LMArena)
1468 rating
11 of 17
4%
Furniture AssemblyEpoch AI
80%
2 of 28
4%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
47.73%
5 of 13
4%
LiveBench LanguageLiveBench
89.43%
4 of 34
3%
LiveBench Instruction FollowingLiveBench
75.58%
7 of 34
3%
LiveBench CodingLiveBench
80.36%
16 of 34
3%
LiveBench Agentic CodingLiveBench
57.32%
13 of 34
3%
ArXivMathMathArena (ETH Zürich)
92.68%
2 of 11
3%
BrokenArXivMathArena (ETH Zürich)
90.67%
3 of 11
3%
FrontierMath Tiers 1–3Epoch AI
93.68%
1 of 34
2%
Mystery Game PuzzlesEpoch AI
84%
1 of 32
2%
LiveBench MathematicsLiveBench
96.81%
4 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
81.98%
3 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
83.08%
1 of 14
1%
GPQA DiamondEpoch AI
95.77%
2 of 32
1%
Chess PuzzlesEpoch AI
74%
1 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Opus 5
Place#5of 33 models with enough tests
Silicon score92.390.3 to 94.3 with other tests
Price$5 / $25in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing83.4
Coding87.6
Math & reasoning91.9
AI agents86.1
Web research93.5
Understanding images98.6
Its tests
Claude Opus 5 took 20 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
59.9%
13 of 32
11%
Arena SearchArena (LMArena)
1242 rating
6 of 17
11%
Arena AgentArena (LMArena)
+9.23 points
8 of 34
9%
τ³-bench Bankingτ³-bench (Sierra)
48.71%
2 of 12
9%
Arena WebDevArena (LMArena)
1690 rating
6 of 37
7%
Arena TextArena (LMArena)
1490 rating
9 of 34
6%
LiveBench ReasoningLiveBench
91.21%
7 of 34
5%
Arena VisionArena (LMArena)
1287 rating
10 of 29
5%
Arena DocumentArena (LMArena)
1516 rating
1 of 17
5%
Furniture AssemblyEpoch AI
60.83%
7 of 28
5%
LiveBench CodingLiveBench
81.45%
11 of 34
4%
LiveBench Agentic CodingLiveBench
65.2%
4 of 34
4%
EBR-benchEpoch AI
45.71%
6 of 17
4%
LiveBench LanguageLiveBench
88.69%
5 of 34
3%
LiveBench Instruction FollowingLiveBench
63.77%
29 of 34
3%
FrontierMath Tiers 1–3Epoch AI
85.61%
11 of 34
2%
Mystery Game PuzzlesEpoch AI
59%
5 of 32
2%
LiveBench MathematicsLiveBench
95.73%
11 of 34
2%
GPQA DiamondEpoch AI
93.88%
10 of 32
1%
Chess PuzzlesEpoch AI
43%
22 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Fable 5
Place#6of 33 models with enough tests
Silicon score91.689.6 to 93.6 with other tests
Price$10 / $50in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing92.0
Coding87.0
Math & reasoning87.6
AI agents80.6
Web research93.1
Understanding images95.2
Its tests
Claude Fable 5 took 21 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
70.7%
7 of 32
11%
Arena SearchArena (LMArena)
1229 rating
10 of 17
10%
Arena AgentArena (LMArena)
+8.49 points
9 of 34
8%
τ³-bench Bankingτ³-bench (Sierra)
39.69%
7 of 12
8%
Arena WebDevArena (LMArena)
1624 rating
16 of 37
7%
MirrorCodeEpoch AI
63.89%
3 of 8
7%
Arena TextArena (LMArena)
1503 rating
3 of 34
6%
Arena VisionArena (LMArena)
1308 rating
1 of 29
5%
Arena DocumentArena (LMArena)
1496 rating
4 of 17
5%
Furniture AssemblyEpoch AI
35.83%
17 of 28
5%
LiveBench ReasoningLiveBench
89.65%
11 of 34
4%
EBR-benchEpoch AI
39.52%
9 of 17
4%
LiveBench LanguageLiveBench
90.68%
1 of 34
3%
LiveBench Instruction FollowingLiveBench
75.77%
6 of 34
3%
LiveBench CodingLiveBench
85.99%
4 of 34
3%
LiveBench Agentic CodingLiveBench
62.17%
7 of 34
3%
FrontierMath Tiers 1–3Epoch AI
87.02%
9 of 34
2%
Mystery Game PuzzlesEpoch AI
52%
10 of 32
2%
LiveBench MathematicsLiveBench
95.99%
8 of 34
2%
GPQA DiamondEpoch AI
85.86%
32 of 32
1%
Chess PuzzlesEpoch AI
47%
16 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-5.6 Sol
Place#7of 33 models with enough tests
Silicon score91.489.5 to 93.3 with other tests
Price$4 / $20in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing85.3
Coding78.9
Math & reasoning93.7
AI agents84.0
Web research100.0
Understanding images92.7
Its tests
GPT-5.6 Sol took 23 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
69.7%
8 of 32
10%
Arena SearchArena (LMArena)
1261 rating
1 of 17
9%
Arena AgentArena (LMArena)
+7.74 points
10 of 34
7%
τ³-bench Bankingτ³-bench (Sierra)
46.91%
4 of 12
7%
Arena WebDevArena (LMArena)
1617 rating
20 of 37
6%
MirrorCodeEpoch AI
20%
6 of 8
6%
Arena TextArena (LMArena)
1485 rating
13 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
88.48%
2 of 13
5%
Arena VisionArena (LMArena)
1287 rating
11 of 29
5%
Arena DocumentArena (LMArena)
1483 rating
7 of 17
5%
Furniture AssemblyEpoch AI
56.67%
9 of 28
5%
LiveBench ReasoningLiveBench
91.65%
5 of 34
4%
EBR-benchEpoch AI
44.76%
7 of 17
4%
LiveBench LanguageLiveBench
87.68%
7 of 34
3%
LiveBench Instruction FollowingLiveBench
71.85%
14 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
68.75%
11 of 15
3%
LiveBench CodingLiveBench
83.94%
5 of 34
3%
LiveBench Agentic CodingLiveBench
56.21%
20 of 34
3%
FrontierMath Tiers 1–3Epoch AI
89.12%
6 of 34
2%
Mystery Game PuzzlesEpoch AI
58%
6 of 32
2%
LiveBench MathematicsLiveBench
96.2%
7 of 34
2%
GPQA DiamondEpoch AI
93.5%
11 of 32
1%
Chess PuzzlesEpoch AI
59%
6 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Sonnet 5.5
Place#8of 33 models with enough tests
Silicon score90.688.3 to 92.9 with other tests
Price$2 / $10in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing77.9
Coding95.0
Math & reasoning95.2
Understanding images88.7
Its tests
Claude Sonnet 5.5 took 16 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
46.5%
23 of 32
15%
Arena VisionArena (LMArena)
1266 rating
21 of 29
14%
Arena AgentArena (LMArena)
+13 points
6 of 34
11%
Arena WebDevArena (LMArena)
1774 rating
3 of 37
9%
Arena TextArena (LMArena)
1475 rating
22 of 34
8%
Furniture AssemblyEpoch AI
75%
4 of 28
7%
LiveBench ReasoningLiveBench
91.63%
6 of 34
6%
LiveBench CodingLiveBench
91.37%
1 of 34
5%
LiveBench Agentic CodingLiveBench
56.31%
19 of 34
5%
LiveBench LanguageLiveBench
83.43%
15 of 34
4%
LiveBench Instruction FollowingLiveBench
70.52%
18 of 34
4%
FrontierMath Tiers 1–3Epoch AI
88.77%
7 of 34
3%
Mystery Game PuzzlesEpoch AI
65%
4 of 32
3%
LiveBench MathematicsLiveBench
96.69%
5 of 34
3%
GPQA DiamondEpoch AI
95.58%
3 of 32
2%
Chess PuzzlesEpoch AI
59%
6 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Alibaba (Qwen)Closed model
Qwen3.8 Max
Place#9of 33 models with enough tests
Silicon score88.186 to 90.2 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 6 of 6
Chat & writing79.2
Coding81.6
Math & reasoning84.3
AI agents83.8
Web research95.2
Understanding images92.5
Its tests
Qwen3.8 Max took 20 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
47.3%
21 of 32
11%
Arena SearchArena (LMArena)
1261 rating
2 of 17
10%
Arena VisionArena (LMArena)
1301 rating
2 of 29
10%
Arena AgentArena (LMArena)
+1.98 points
23 of 34
8%
τ³-bench Bankingτ³-bench (Sierra)
55.15%
1 of 12
8%
Arena WebDevArena (LMArena)
1673 rating
9 of 37
7%
Arena TextArena (LMArena)
1482 rating
15 of 34
6%
LiveBench ReasoningLiveBench
88.21%
19 of 34
5%
Furniture AssemblyEpoch AI
20%
28 of 28
5%
LiveBench LanguageLiveBench
79.69%
24 of 34
3%
LiveBench Instruction FollowingLiveBench
74.08%
10 of 34
3%
LiveBench CodingLiveBench
72.87%
31 of 34
3%
LiveBench Agentic CodingLiveBench
64.65%
5 of 34
3%
ArXivMathMathArena (ETH Zürich)
67.59%
5 of 11
3%
BrokenArXivMathArena (ETH Zürich)
48.71%
7 of 11
3%
FrontierMath Tiers 1–3Epoch AI
74.74%
17 of 34
2%
Mystery Game PuzzlesEpoch AI
38%
14 of 32
2%
LiveBench MathematicsLiveBench
91.31%
23 of 34
2%
GPQA DiamondEpoch AI
92.68%
16 of 32
1%
Chess PuzzlesEpoch AI
46%
19 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-6 Sol
Place#10of 33 models with enough tests
Silicon score87.585.6 to 89.4 with other tests
Price$2 / $10in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing80.4
Coding83.0
Math & reasoning94.1
AI agents89.1
Web research81.7
Understanding images88.1
Its tests
GPT-6 Sol took 24 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
60.7%
12 of 32
10%
Arena SearchArena (LMArena)
1175 rating
16 of 17
9%
Arena VisionArena (LMArena)
1270 rating
19 of 29
9%
Arena AgentArena (LMArena)
+10.51 points
7 of 34
8%
Arena WebDevArena (LMArena)
1686 rating
7 of 37
6%
Arena TextArena (LMArena)
1458 rating
30 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
86.15%
5 of 13
5%
Furniture AssemblyEpoch AI
58.33%
8 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
43.06%
7 of 13
5%
LiveBench ReasoningLiveBench
88.65%
18 of 34
4%
EBR-benchEpoch AI
53.33%
5 of 17
4%
LiveBench LanguageLiveBench
85.3%
13 of 34
3%
LiveBench Instruction FollowingLiveBench
68.57%
21 of 34
3%
LiveBench CodingLiveBench
81.77%
10 of 34
3%
LiveBench Agentic CodingLiveBench
52.88%
26 of 34
3%
ArXivMathMathArena (ETH Zürich)
91.32%
3 of 11
3%
BrokenArXivMathArena (ETH Zürich)
86.98%
4 of 11
3%
FrontierMath Tiers 1–3Epoch AI
89.82%
5 of 34
2%
Mystery Game PuzzlesEpoch AI
56%
8 of 32
2%
LiveBench MathematicsLiveBench
96.36%
6 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
85.22%
1 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
75.59%
7 of 14
1%
GPQA DiamondEpoch AI
94.32%
8 of 32
1%
Chess PuzzlesEpoch AI
56%
10 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
GoogleClosed model
Gemini 3.7 Flash
Place#11of 33 models with enough tests
Silicon score87.585.4 to 89.6 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 5 of 6
Chat & writing88.4
Coding77.5
Math & reasoning84.8
Web research87.7
Understanding images94.6
Its tests
Gemini 3.7 Flash took 20 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
69.2%
10 of 32
13%
Arena VisionArena (LMArena)
1297 rating
5 of 29
12%
Arena AgentArena (LMArena)
-0.52 points
30 of 34
9%
Arena WebDevArena (LMArena)
1590 rating
23 of 37
8%
Arena TextArena (LMArena)
1487 rating
11 of 34
7%
FACTS SearchFACTS (Google DeepMind and Kaggle)
80.89%
9 of 13
6%
Furniture AssemblyEpoch AI
26.67%
23 of 28
6%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
49.83%
2 of 13
6%
LiveBench ReasoningLiveBench
87.8%
20 of 34
5%
LiveBench CodingLiveBench
78.89%
22 of 34
4%
LiveBench Agentic CodingLiveBench
58.28%
12 of 34
4%
LiveBench LanguageLiveBench
85.46%
11 of 34
3%
LiveBench Instruction FollowingLiveBench
79.93%
2 of 34
3%
FrontierMath Tiers 1–3Epoch AI
71.58%
20 of 34
3%
Mystery Game PuzzlesEpoch AI
37%
15 of 32
3%
LiveBench MathematicsLiveBench
93.47%
16 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
75.54%
6 of 15
2%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
78.72%
4 of 14
2%
GPQA DiamondEpoch AI
94.82%
6 of 32
1%
Chess PuzzlesEpoch AI
47%
16 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-5.5
Place#12of 33 models with enough tests
Silicon score87.385.4 to 89.2 with other tests
Price$5 / $30in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing84.7
Coding72.2
Math & reasoning90.3
AI agents78.8
Web research95.4
Understanding images91.8
Its tests
GPT-5.5 took 24 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
63%
11 of 32
10%
Arena SearchArena (LMArena)
1247 rating
4 of 17
9%
Arena AgentArena (LMArena)
+3.48 points
16 of 34
7%
τ³-bench Bankingτ³-bench (Sierra)
44.59%
5 of 12
7%
Arena WebDevArena (LMArena)
1513 rating
35 of 37
6%
MirrorCodeEpoch AI
10%
7 of 8
6%
Arena TextArena (LMArena)
1481 rating
17 of 34
5%
Arena VisionArena (LMArena)
1286 rating
14 of 29
5%
Arena DocumentArena (LMArena)
1486 rating
6 of 17
5%
Furniture AssemblyEpoch AI
44.17%
12 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
45.77%
6 of 13
5%
LiveBench ReasoningLiveBench
89.65%
11 of 34
4%
EBR-benchEpoch AI
34.29%
10 of 17
4%
LiveBench LanguageLiveBench
87.36%
8 of 34
3%
LiveBench Instruction FollowingLiveBench
70.73%
17 of 34
3%
LiveBench CodingLiveBench
82.15%
7 of 34
3%
LiveBench Agentic CodingLiveBench
53.99%
24 of 34
3%
FrontierMath Tiers 1–3Epoch AI
85.26%
12 of 34
2%
Mystery Game PuzzlesEpoch AI
56%
8 of 32
2%
LiveBench MathematicsLiveBench
95.86%
10 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
76.19%
5 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
78.04%
5 of 14
1%
GPQA DiamondEpoch AI
90.66%
23 of 32
1%
Chess PuzzlesEpoch AI
63%
4 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
MetaClosed model
Muse Spark 1.3
Place#13of 33 models with enough tests
Silicon score87.185 to 89.2 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 3 of 6
Chat & writing87.4
Coding85.0
Math & reasoning83.8
Its tests
Muse Spark 1.3 took 16 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
Arena AgentArena (LMArena)
+3.5 points
15 of 34
13%
Arena WebDevArena (LMArena)
1659 rating
10 of 37
11%
Arena TextArena (LMArena)
1494 rating
8 of 34
9%
Arena VisionArena (LMArena)
1290 rating
8 of 29
8%
Arena DocumentArena (LMArena)
1471 rating
10 of 17
8%
LiveBench ReasoningLiveBench
89.65%
11 of 34
7%
LiveBench LanguageLiveBench
82.79%
19 of 34
5%
LiveBench Instruction FollowingLiveBench
78%
4 of 34
5%
LiveBench CodingLiveBench
81.06%
13 of 34
5%
LiveBench Agentic CodingLiveBench
64.09%
6 of 34
5%
ArXivMathMathArena (ETH Zürich)
65.23%
6 of 11
5%
BrokenArXivMathArena (ETH Zürich)
36.66%
10 of 11
5%
FrontierMath Tiers 1–3Epoch AI
74.39%
18 of 34
4%
Mystery Game PuzzlesEpoch AI
25%
26 of 32
4%
LiveBench MathematicsLiveBench
95.95%
9 of 34
4%
Chess PuzzlesEpoch AI
37%
26 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
GoogleClosed model
Gemini 3.8 Flash
Place#14of 33 models with enough tests
Silicon score86.584.5 to 88.5 with other tests
Price$0.75 / $3.75in / out, per million tokens, until December 31. Then $1.50 / $7.50.
Each job on its own · scored in 5 of 6
Chat & writing90.5
Coding72.1
Math & reasoning83.0
Web research82.3
Understanding images93.4
Its tests
Gemini 3.8 Flash took 22 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
69.7%
8 of 32
12%
Arena VisionArena (LMArena)
1289 rating
9 of 29
11%
Arena AgentArena (LMArena)
+4.02 points
13 of 34
9%
Arena WebDevArena (LMArena)
1585 rating
25 of 37
7%
Arena TextArena (LMArena)
1496 rating
7 of 34
6%
FACTS SearchFACTS (Google DeepMind and Kaggle)
76.33%
13 of 13
5%
LiveBench ReasoningLiveBench
89.29%
14 of 34
5%
Furniture AssemblyEpoch AI
31.67%
21 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
50.7%
1 of 13
5%
LiveBench CodingLiveBench
72.49%
33 of 34
4%
LiveBench Agentic CodingLiveBench
54.24%
23 of 34
4%
ArXivMathMathArena (ETH Zürich)
55.35%
10 of 11
4%
BrokenArXivMathArena (ETH Zürich)
22.4%
11 of 11
4%
LiveBench LanguageLiveBench
87.79%
6 of 34
3%
LiveBench Instruction FollowingLiveBench
81.41%
1 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
73.09%
8 of 15
2%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
77.34%
6 of 14
2%
FrontierMath Tiers 1–3Epoch AI
68.42%
23 of 34
2%
Mystery Game PuzzlesEpoch AI
47%
11 of 32
2%
LiveBench MathematicsLiveBench
91.56%
22 of 34
2%
GPQA DiamondEpoch AI
95.39%
4 of 32
1%
Chess PuzzlesEpoch AI
58%
9 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Opus 4.7
Place#15of 33 models with enough tests
Silicon score83.982 to 85.8 with other tests
Price$5 / $25in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing79.6
Coding75.9
Math & reasoning81.0
AI agents74.1
Web research88.9
Understanding images88.9
Its tests
Claude Opus 4.7 took 23 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
51.7%
17 of 32
11%
Arena SearchArena (LMArena)
1230 rating
9 of 17
10%
τ³-bench Bankingτ³-bench (Sierra)
40.21%
6 of 12
8%
Arena WebDevArena (LMArena)
1558 rating
28 of 37
7%
MirrorCodeEpoch AI
31.11%
5 of 8
7%
Arena TextArena (LMArena)
1501 rating
5 of 34
6%
Arena VisionArena (LMArena)
1298 rating
4 of 29
5%
Arena DocumentArena (LMArena)
1495 rating
5 of 17
5%
Furniture AssemblyEpoch AI
33.33%
20 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
35.84%
11 of 13
5%
LiveBench ReasoningLiveBench
87.19%
22 of 34
4%
EBR-benchEpoch AI
19.05%
13 of 17
4%
LiveBench LanguageLiveBench
77.91%
26 of 34
3%
LiveBench Instruction FollowingLiveBench
66.74%
24 of 34
3%
LiveBench CodingLiveBench
82.09%
8 of 34
3%
LiveBench Agentic CodingLiveBench
50.66%
29 of 34
3%
FrontierMath Tiers 1–3Epoch AI
70.18%
21 of 34
2%
Mystery Game PuzzlesEpoch AI
28%
24 of 32
2%
LiveBench MathematicsLiveBench
92.85%
19 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
61.73%
14 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
38.82%
12 of 14
1%
GPQA DiamondEpoch AI
90.15%
28 of 32
1%
Chess PuzzlesEpoch AI
41%
24 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Moonshot AI (Kimi)Open model
Kimi K3
Place#16of 33 models with enough tests
Silicon score83.881.6 to 86 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 4 of 6
Chat & writing81.9
Coding84.4
Math & reasoning81.8
AI agents72.7
Its tests
Kimi K3 took 18 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
50.6%
18 of 32
14%
Arena AgentArena (LMArena)
+1.9 points
24 of 34
10%
τ³-bench Bankingτ³-bench (Sierra)
37.11%
9 of 12
10%
Arena WebDevArena (LMArena)
1655 rating
11 of 37
9%
Arena TextArena (LMArena)
1488 rating
10 of 34
7%
Furniture AssemblyEpoch AI
34.17%
18 of 28
7%
LiveBench ReasoningLiveBench
90.67%
8 of 34
6%
LiveBench LanguageLiveBench
85.53%
10 of 34
4%
LiveBench Instruction FollowingLiveBench
71.36%
16 of 34
4%
LiveBench CodingLiveBench
81.45%
11 of 34
4%
LiveBench Agentic CodingLiveBench
62.17%
7 of 34
4%
ArXivMathMathArena (ETH Zürich)
57.26%
8 of 11
4%
BrokenArXivMathArena (ETH Zürich)
52.09%
6 of 11
4%
FrontierMath Tiers 1–3Epoch AI
72.18%
19 of 34
3%
Mystery Game PuzzlesEpoch AI
26%
25 of 32
3%
LiveBench MathematicsLiveBench
84.44%
32 of 34
3%
GPQA DiamondEpoch AI
93.12%
14 of 32
1%
Chess PuzzlesEpoch AI
37%
26 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Opus 4.8
Place#17of 33 models with enough tests
Silicon score83.481.5 to 85.3 with other tests
Price$5 / $25in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing78.6
Coding74.2
Math & reasoning86.3
AI agents78.8
Web research81.5
Understanding images87.7
Its tests
Claude Opus 4.8 took 24 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
53%
15 of 32
10%
Arena SearchArena (LMArena)
1204 rating
14 of 17
9%
Arena AgentArena (LMArena)
+7.55 points
11 of 34
8%
τ³-bench Bankingτ³-bench (Sierra)
39.69%
7 of 12
8%
Arena WebDevArena (LMArena)
1556 rating
29 of 37
6%
Arena TextArena (LMArena)
1482 rating
16 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
79.6%
10 of 13
5%
Arena VisionArena (LMArena)
1286 rating
15 of 29
5%
Arena DocumentArena (LMArena)
1475 rating
8 of 17
5%
Furniture AssemblyEpoch AI
42.5%
14 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
39.77%
9 of 13
5%
LiveBench ReasoningLiveBench
89.19%
15 of 34
4%
EBR-benchEpoch AI
28.57%
12 of 17
4%
LiveBench LanguageLiveBench
79.66%
25 of 34
3%
LiveBench Instruction FollowingLiveBench
72.03%
12 of 34
3%
LiveBench CodingLiveBench
81.83%
9 of 34
3%
LiveBench Agentic CodingLiveBench
50.51%
30 of 34
3%
FrontierMath Tiers 1–3Epoch AI
80%
14 of 34
2%
Mystery Game PuzzlesEpoch AI
36%
16 of 32
2%
LiveBench MathematicsLiveBench
94.32%
15 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
69.68%
10 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
25.06%
14 of 14
1%
GPQA DiamondEpoch AI
91.04%
20 of 32
1%
Chess PuzzlesEpoch AI
52%
13 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Opus 4.6
Place#18of 33 models with enough tests
Silicon score82.280.2 to 84.2 with other tests
Price$5 / $25in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing78.7
Coding71.0
Math & reasoning78.9
AI agents62.6
Web research92.8
Understanding images90.2
Its tests
Claude Opus 4.6 took 22 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
47%
22 of 32
11%
Arena SearchArena (LMArena)
1251 rating
3 of 17
11%
τ³-bench Bankingτ³-bench (Sierra)
27.32%
11 of 12
9%
Arena WebDevArena (LMArena)
1546 rating
31 of 37
7%
Arena TextArena (LMArena)
1503 rating
4 of 34
6%
LiveBench ReasoningLiveBench
88.67%
17 of 34
5%
Arena VisionArena (LMArena)
1299 rating
3 of 29
5%
Arena DocumentArena (LMArena)
1507 rating
3 of 17
5%
Furniture AssemblyEpoch AI
28.33%
22 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
36.75%
10 of 13
5%
LiveBench CodingLiveBench
78.18%
25 of 34
4%
LiveBench Agentic CodingLiveBench
48.99%
31 of 34
4%
EBR-benchEpoch AI
12.7%
15 of 17
4%
LiveBench LanguageLiveBench
83.27%
16 of 34
3%
LiveBench Instruction FollowingLiveBench
63.31%
30 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
54.91%
15 of 15
2%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
32.95%
13 of 14
2%
FrontierMath Tiers 1–3Epoch AI
65.96%
25 of 34
2%
Mystery Game PuzzlesEpoch AI
25%
26 of 32
2%
LiveBench MathematicsLiveBench
89.32%
27 of 34
2%
GPQA DiamondEpoch AI
90.53%
25 of 32
1%
Chess PuzzlesEpoch AI
21%
33 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
DeepSeekOpen model
DeepSeek V4.1 Flash
Place#19of 33 models with enough tests
Silicon score81.979.7 to 84.1 with other tests
Price$0.30 / $1.20in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing76.1
Coding86.8
Math & reasoning80.9
Understanding images79.5
Its tests
DeepSeek V4.1 Flash took 18 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
42.9%
25 of 32
14%
Arena VisionArena (LMArena)
1260 rating
25 of 29
13%
Arena AgentArena (LMArena)
+2.94 points
19 of 34
10%
Arena WebDevArena (LMArena)
1619 rating
18 of 37
8%
Arena TextArena (LMArena)
1475 rating
23 of 34
7%
LiveBench ReasoningLiveBench
86.69%
24 of 34
6%
Furniture AssemblyEpoch AI
34.17%
18 of 28
6%
LiveBench LanguageLiveBench
81.19%
21 of 34
4%
LiveBench Instruction FollowingLiveBench
70.03%
19 of 34
4%
LiveBench CodingLiveBench
80.04%
17 of 34
4%
LiveBench Agentic CodingLiveBench
77.27%
1 of 34
4%
ArXivMathMathArena (ETH Zürich)
53.28%
11 of 11
4%
BrokenArXivMathArena (ETH Zürich)
40.87%
9 of 11
4%
FrontierMath Tiers 1–3Epoch AI
67.37%
24 of 34
3%
Mystery Game PuzzlesEpoch AI
43%
12 of 32
3%
LiveBench MathematicsLiveBench
93.29%
17 of 34
3%
GPQA DiamondEpoch AI
89.84%
30 of 32
1%
Chess PuzzlesEpoch AI
48%
14 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-5.6 Terra
Place#20of 33 models with enough tests
Silicon score81.379.3 to 83.3 with other tests
Price$2 / $12in / out, per million tokens
Each job on its own · scored in 5 of 6
Chat & writing74.6
Coding71.7
Math & reasoning89.0
Web research79.6
Understanding images87.0
Its tests
GPT-5.6 Terra took 20 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
43.2%
24 of 32
13%
Arena AgentArena (LMArena)
+0.54 points
27 of 34
10%
Arena WebDevArena (LMArena)
1522 rating
33 of 37
8%
Arena TextArena (LMArena)
1466 rating
25 of 34
7%
FACTS SearchFACTS (Google DeepMind and Kaggle)
81.74%
7 of 13
6%
LiveBench ReasoningLiveBench
90.63%
9 of 34
6%
Arena VisionArena (LMArena)
1268 rating
20 of 29
6%
Arena DocumentArena (LMArena)
1472 rating
9 of 17
6%
Furniture AssemblyEpoch AI
54.17%
10 of 28
6%
LiveBench LanguageLiveBench
82.89%
17 of 34
4%
LiveBench Instruction FollowingLiveBench
64.62%
26 of 34
4%
LiveBench CodingLiveBench
78.25%
24 of 34
4%
LiveBench Agentic CodingLiveBench
54.95%
21 of 34
4%
FrontierMath Tiers 1–3Epoch AI
85.96%
10 of 34
3%
Mystery Game PuzzlesEpoch AI
35%
17 of 32
3%
LiveBench MathematicsLiveBench
94.91%
14 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
71.21%
9 of 15
2%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
55.04%
10 of 14
2%
GPQA DiamondEpoch AI
93.31%
13 of 32
1%
Chess PuzzlesEpoch AI
47%
16 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
xAIClosed model
Grok 4.6
Place#21of 33 models with enough tests
Silicon score80.978.8 to 83 with other tests
Price$2 / $6in / out, per million tokens
Each job on its own · scored in 5 of 6
Chat & writing75.3
Coding77.5
Math & reasoning84.5
AI agents74.0
Understanding images80.5
Its tests
Grok 4.6 took 18 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
49.3%
19 of 32
14%
Arena AgentArena (LMArena)
+1.18 points
26 of 34
11%
Arena WebDevArena (LMArena)
1617 rating
19 of 37
9%
Arena TextArena (LMArena)
1455 rating
31 of 34
8%
Arena VisionArena (LMArena)
1262 rating
24 of 29
7%
Arena DocumentArena (LMArena)
1452 rating
14 of 17
7%
Furniture AssemblyEpoch AI
40%
16 of 28
7%
LiveBench ReasoningLiveBench
90.51%
10 of 34
6%
EBR-benchEpoch AI
30.48%
11 of 17
5%
LiveBench LanguageLiveBench
83.7%
14 of 34
4%
LiveBench Instruction FollowingLiveBench
71.87%
13 of 34
4%
LiveBench CodingLiveBench
76.78%
29 of 34
4%
LiveBench Agentic CodingLiveBench
57.02%
15 of 34
4%
FrontierMath Tiers 1–3Epoch AI
65.96%
25 of 34
3%
Mystery Game PuzzlesEpoch AI
34%
19 of 32
3%
LiveBench MathematicsLiveBench
92.57%
21 of 34
3%
GPQA DiamondEpoch AI
94%
9 of 32
1%
Chess PuzzlesEpoch AI
33%
31 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
DeepSeekOpen model
DeepSeek V4 Pro 0813
Place#22of 33 models with enough tests
Silicon score80.878.3 to 83.3 with other tests
Price$1.32 / $3.96in / out, per million tokens
Each job on its own · scored in 3 of 6
Chat & writing76.3
Coding74.8
Math & reasoning82.1
Its tests
DeepSeek V4 Pro 0813 took 14 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
52.91%
16 of 32
19%
Arena AgentArena (LMArena)
+0.37 points
28 of 34
14%
Arena WebDevArena (LMArena)
1582 rating
26 of 37
12%
Arena TextArena (LMArena)
1464 rating
27 of 34
10%
LiveBench ReasoningLiveBench
85.84%
25 of 34
8%
LiveBench CodingLiveBench
77.16%
27 of 34
6%
LiveBench Agentic CodingLiveBench
54.95%
21 of 34
6%
LiveBench LanguageLiveBench
82.07%
20 of 34
5%
LiveBench Instruction FollowingLiveBench
67.7%
22 of 34
5%
FrontierMath Tiers 1–3Epoch AI
64.56%
28 of 34
4%
Mystery Game PuzzlesEpoch AI
43%
12 of 32
4%
LiveBench MathematicsLiveBench
95.09%
13 of 34
4%
GPQA DiamondEpoch AI
91.67%
18 of 32
2%
Chess PuzzlesEpoch AI
46%
19 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Z.ai (GLM)Open model
GLM-5.3
Place#23of 33 models with enough tests
Silicon score80.778.2 to 83.2 with other tests
Price$1.40 / $4.40in / out, per million tokens
Each job on its own · scored in 3 of 6
Chat & writing75.9
Coding80.4
Math & reasoning78.1
Its tests
GLM-5.3 took 14 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
41%
27 of 32
19%
Arena AgentArena (LMArena)
+2.09 points
22 of 34
14%
Arena WebDevArena (LMArena)
1621 rating
17 of 37
12%
Arena TextArena (LMArena)
1479 rating
19 of 34
10%
LiveBench ReasoningLiveBench
85.8%
26 of 34
8%
LiveBench CodingLiveBench
78.95%
19 of 34
6%
LiveBench Agentic CodingLiveBench
60.91%
10 of 34
6%
LiveBench LanguageLiveBench
79.86%
23 of 34
5%
LiveBench Instruction FollowingLiveBench
69.3%
20 of 34
5%
FrontierMath Tiers 1–3Epoch AI
68.77%
22 of 34
4%
Mystery Game PuzzlesEpoch AI
33%
21 of 32
4%
LiveBench MathematicsLiveBench
87.9%
29 of 34
4%
GPQA DiamondEpoch AI
90.91%
21 of 32
2%
Chess PuzzlesEpoch AI
37%
26 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
xAIClosed model
Grok 4.5
Place#24of 33 models with enough tests
Silicon score80.378.3 to 82.3 with other tests
Price$2 / $6in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing77.1
Coding68.8
Math & reasoning78.9
AI agents77.8
Web research82.7
Understanding images80.9
Its tests
Grok 4.5 took 21 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
48.3%
20 of 32
11%
Arena SearchArena (LMArena)
1216 rating
11 of 17
10%
Arena AgentArena (LMArena)
+0.28 points
29 of 34
8%
τ³-bench Bankingτ³-bench (Sierra)
47.94%
3 of 12
8%
Arena WebDevArena (LMArena)
1553 rating
30 of 37
7%
Arena TextArena (LMArena)
1466 rating
26 of 34
6%
FACTS SearchFACTS (Google DeepMind and Kaggle)
79.44%
11 of 13
5%
LiveBench ReasoningLiveBench
87.17%
23 of 34
5%
Arena VisionArena (LMArena)
1280 rating
16 of 29
5%
Arena DocumentArena (LMArena)
1452 rating
15 of 17
5%
Furniture AssemblyEpoch AI
22.5%
25 of 28
5%
LiveBench LanguageLiveBench
82.8%
18 of 34
3%
LiveBench Instruction FollowingLiveBench
71.53%
15 of 34
3%
LiveBench CodingLiveBench
68.59%
34 of 34
3%
LiveBench Agentic CodingLiveBench
56.46%
18 of 34
3%
FrontierMath Tiers 1–3Epoch AI
57.19%
31 of 34
3%
GPQA DiamondEpoch AI
93.43%
12 of 32
2%
Chess PuzzlesEpoch AI
35%
30 of 35
2%
LiveBench MathematicsLiveBench
90.82%
25 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
74.52%
7 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
46.83%
11 of 14
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
GoogleClosed model
Gemini 3.1 Pro
Place#25of 33 models with enough tests
Silicon score79.978.1 to 81.7 with other tests
Price$2 / $12in / out, per million tokens
Each job on its own · scored in 6 of 6
Chat & writing88.1
Coding64.8
Math & reasoning79.0
AI agents62.7
Web research88.5
Understanding images82.0
Its tests
Gemini 3.1 Pro took 25 of our tests, from 5 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
73.5%
4 of 32
9%
Arena SearchArena (LMArena)
1209 rating
13 of 17
9%
Arena AgentArena (LMArena)
-3.41 points
34 of 34
7%
τ³-bench Bankingτ³-bench (Sierra)
26.03%
12 of 12
7%
Arena WebDevArena (LMArena)
1447 rating
37 of 37
6%
MirrorCodeEpoch AI
8.89%
8 of 8
6%
Arena TextArena (LMArena)
1487 rating
12 of 34
5%
FACTS SearchFACTS (Google DeepMind and Kaggle)
83.7%
6 of 13
4%
LiveBench ReasoningLiveBench
84%
29 of 34
4%
EBR-benchEpoch AI
14.29%
14 of 17
4%
Arena VisionArena (LMArena)
1280 rating
17 of 29
4%
Arena DocumentArena (LMArena)
1444 rating
17 of 17
4%
Furniture AssemblyEpoch AI
26.67%
23 of 28
4%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
41.34%
8 of 13
4%
LiveBench LanguageLiveBench
85.38%
12 of 34
3%
LiveBench Instruction FollowingLiveBench
79.1%
3 of 34
3%
LiveBench CodingLiveBench
76.45%
30 of 34
3%
LiveBench Agentic CodingLiveBench
44.14%
34 of 34
3%
FrontierMath Tiers 1–3Epoch AI
59.65%
29 of 34
2%
Mystery Game PuzzlesEpoch AI
34%
19 of 32
2%
LiveBench MathematicsLiveBench
91.04%
24 of 34
2%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
65.67%
13 of 15
1%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
78.96%
3 of 14
1%
GPQA DiamondEpoch AI
94.44%
7 of 32
1%
Chess PuzzlesEpoch AI
55%
11 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Alibaba (Qwen)Open modelFew tests yet
Qwen3.8 Flash Next
Place–None yet: few tests yet
Silicon score79.476.4 to 82.4 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 1 of 6
Coding77.8
Its tests
Qwen3.8 Flash Next took 8 of our tests, from 2 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
Arena AgentArena (LMArena)
-2.55 points
33 of 34
24%
Arena WebDevArena (LMArena)
1633 rating
13 of 37
20%
LiveBench ReasoningLiveBench
87.38%
21 of 34
13%
LiveBench CodingLiveBench
72.55%
32 of 34
10%
LiveBench Agentic CodingLiveBench
61.62%
9 of 34
10%
LiveBench LanguageLiveBench
74.64%
31 of 34
8%
LiveBench Instruction FollowingLiveBench
77.11%
5 of 34
8%
LiveBench MathematicsLiveBench
85.82%
31 of 34
7%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it. Few tests yet: it has a job score of its own in only 1 of the 6 jobs, so the score may still move. It is listed, but gets no place.
xAIClosed model
Grok 4.7
Place#26of 33 models with enough tests
Silicon score78.776.5 to 80.9 with other tests
Price$2 / $6in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing74.7
Coding78.0
Math & reasoning78.3
Understanding images72.3
Its tests
Grok 4.7 took 18 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
56%
14 of 32
14%
Arena VisionArena (LMArena)
1248 rating
27 of 29
13%
Arena AgentArena (LMArena)
+3.33 points
17 of 34
10%
Arena WebDevArena (LMArena)
1640 rating
12 of 37
8%
Arena TextArena (LMArena)
1444 rating
33 of 34
7%
LiveBench ReasoningLiveBench
82.65%
30 of 34
6%
Furniture AssemblyEpoch AI
20.83%
27 of 28
6%
LiveBench LanguageLiveBench
80.14%
22 of 34
4%
LiveBench Instruction FollowingLiveBench
75.28%
8 of 34
4%
LiveBench CodingLiveBench
77.16%
27 of 34
4%
LiveBench Agentic CodingLiveBench
53.99%
24 of 34
4%
ArXivMathMathArena (ETH Zürich)
59.15%
7 of 11
4%
BrokenArXivMathArena (ETH Zürich)
42.63%
8 of 11
4%
FrontierMath Tiers 1–3Epoch AI
52.98%
34 of 34
3%
Mystery Game PuzzlesEpoch AI
29%
23 of 32
3%
LiveBench MathematicsLiveBench
95.68%
12 of 34
3%
GPQA DiamondEpoch AI
92.74%
15 of 32
1%
Chess PuzzlesEpoch AI
41%
24 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Sonnet 5
Place#27of 33 models with enough tests
Silicon score78.376.1 to 80.5 with other tests
Price$2 / $10in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing68.3
Coding75.8
Math & reasoning82.7
Web research75.4
Its tests
Claude Sonnet 5 took 17 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
33.7%
31 of 32
14%
Arena SearchArena (LMArena)
1191 rating
15 of 17
13%
Arena AgentArena (LMArena)
+4.88 points
12 of 34
10%
Arena WebDevArena (LMArena)
1540 rating
32 of 37
9%
Arena TextArena (LMArena)
1461 rating
28 of 34
7%
Arena VisionArena (LMArena)
1263 rating
23 of 29
7%
Arena DocumentArena (LMArena)
1466 rating
12 of 17
7%
LiveBench ReasoningLiveBench
88.69%
16 of 34
6%
LiveBench LanguageLiveBench
74.97%
30 of 34
4%
LiveBench Instruction FollowingLiveBench
63.86%
28 of 34
4%
LiveBench CodingLiveBench
80.68%
15 of 34
4%
LiveBench Agentic CodingLiveBench
59.39%
11 of 34
4%
FrontierMath Tiers 1–3Epoch AI
65.61%
27 of 34
3%
Mystery Game PuzzlesEpoch AI
35%
17 of 32
3%
LiveBench MathematicsLiveBench
92.94%
18 of 34
3%
GPQA DiamondEpoch AI
90.53%
25 of 32
1%
Chess PuzzlesEpoch AI
45%
21 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
AnthropicClosed model
Claude Haiku 5.5
Place#28of 33 models with enough tests
Silicon score77.274.8 to 79.6 with other tests
Price$0.10 / $0.50in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing59.8
Coding76.2
Math & reasoning80.3
AI agents76.5
Its tests
Claude Haiku 5.5 took 15 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
23.8%
32 of 32
18%
Arena AgentArena (LMArena)
+1.48 points
25 of 34
13%
Arena WebDevArena (LMArena)
1585 rating
24 of 37
11%
Furniture AssemblyEpoch AI
47.5%
11 of 28
8%
LiveBench ReasoningLiveBench
84.63%
28 of 34
7%
EBR-benchEpoch AI
43.81%
8 of 17
7%
LiveBench CodingLiveBench
77.86%
26 of 34
6%
LiveBench Agentic CodingLiveBench
57.17%
14 of 34
6%
LiveBench LanguageLiveBench
63.52%
34 of 34
5%
LiveBench Instruction FollowingLiveBench
66.47%
25 of 34
5%
FrontierMath Tiers 1–3Epoch AI
75.09%
16 of 34
4%
Mystery Game PuzzlesEpoch AI
30%
22 of 32
4%
LiveBench MathematicsLiveBench
92.75%
20 of 34
4%
GPQA DiamondEpoch AI
89.65%
31 of 32
2%
Chess PuzzlesEpoch AI
55%
11 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Z.ai (GLM)Open model
GLM-5.2
Place#29of 33 models with enough tests
Silicon score76.374 to 78.6 with other tests
Price$1.40 / $4.40in / out, per million tokens
Each job on its own · scored in 4 of 6
Chat & writing70.9
Coding76.4
Math & reasoning69.6
AI agents72.1
Its tests
GLM-5.2 took 16 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
34.2%
30 of 32
16%
Arena AgentArena (LMArena)
+2.57 points
21 of 34
12%
τ³-bench Bankingτ³-bench (Sierra)
37.11%
9 of 12
12%
Arena WebDevArena (LMArena)
1604 rating
22 of 37
10%
Arena TextArena (LMArena)
1475 rating
21 of 34
8%
LiveBench ReasoningLiveBench
78.63%
33 of 34
6%
EBR-benchEpoch AI
9.52%
16 of 17
6%
LiveBench CodingLiveBench
79.65%
18 of 34
5%
LiveBench Agentic CodingLiveBench
51.77%
27 of 34
5%
LiveBench LanguageLiveBench
76.24%
28 of 34
4%
LiveBench Instruction FollowingLiveBench
62.29%
31 of 34
4%
FrontierMath Tiers 1–3Epoch AI
59.21%
30 of 34
3%
Mystery Game PuzzlesEpoch AI
19%
29 of 32
3%
LiveBench MathematicsLiveBench
89.78%
26 of 34
3%
GPQA DiamondEpoch AI
91.86%
17 of 32
2%
Chess PuzzlesEpoch AI
22%
32 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-5.6 Luna
Place#30of 33 models with enough tests
Silicon score75.673.6 to 77.6 with other tests
Price$0.20 / $1.20in / out, per million tokens
Each job on its own · scored in 5 of 6
Chat & writing66.8
Coding71.7
Math & reasoning79.0
Web research76.0
Understanding images79.9
Its tests
GPT-5.6 Luna took 19 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
41%
27 of 32
13%
Arena AgentArena (LMArena)
-1.76 points
32 of 34
10%
Arena WebDevArena (LMArena)
1518 rating
34 of 37
8%
Arena TextArena (LMArena)
1453 rating
32 of 34
7%
FACTS SearchFACTS (Google DeepMind and Kaggle)
79.35%
12 of 13
6%
Arena VisionArena (LMArena)
1258 rating
26 of 29
6%
Arena DocumentArena (LMArena)
1457 rating
13 of 17
6%
Furniture AssemblyEpoch AI
42.5%
14 of 28
6%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
35.81%
12 of 13
6%
LiveBench ReasoningLiveBench
85.64%
27 of 34
5%
LiveBench CodingLiveBench
82.92%
6 of 34
4%
LiveBench Agentic CodingLiveBench
48.43%
32 of 34
4%
LiveBench LanguageLiveBench
72.57%
33 of 34
3%
LiveBench Instruction FollowingLiveBench
60.12%
32 of 34
3%
FrontierMath Tiers 1–3Epoch AI
82.11%
13 of 34
3%
Mystery Game PuzzlesEpoch AI
21%
28 of 32
3%
LiveBench MathematicsLiveBench
87.2%
30 of 34
3%
GPQA DiamondEpoch AI
91.6%
19 of 32
1%
Chess PuzzlesEpoch AI
43%
22 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Z.ai (GLM)Open model
GLM-5.3 Flash
Place#31of 33 models with enough tests
Silicon score75.372.8 to 77.8 with other tests
Price$0.15 / $0.50in / out, per million tokens
Each job on its own · scored in 3 of 6
Chat & writing70.0
Coding78.0
Math & reasoning63.2
Its tests
GLM-5.3 Flash took 14 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
Arena VisionArena (LMArena)
1273 rating
18 of 29
18%
Arena AgentArena (LMArena)
-1 points
31 of 34
14%
Arena WebDevArena (LMArena)
1606 rating
21 of 37
12%
Arena TextArena (LMArena)
1474 rating
24 of 34
10%
LiveBench ReasoningLiveBench
77.64%
34 of 34
8%
LiveBench CodingLiveBench
78.95%
19 of 34
6%
LiveBench Agentic CodingLiveBench
56.77%
16 of 34
6%
LiveBench LanguageLiveBench
77.31%
27 of 34
5%
LiveBench Instruction FollowingLiveBench
52.82%
34 of 34
5%
FrontierMath Tiers 1–3Epoch AI
55.79%
33 of 34
4%
Mystery Game PuzzlesEpoch AI
8%
31 of 32
4%
LiveBench MathematicsLiveBench
81.24%
34 of 34
4%
GPQA DiamondEpoch AI
90.15%
28 of 32
2%
Chess PuzzlesEpoch AI
21%
33 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
OpenAIClosed model
GPT-6 Luna
Place#32of 33 models with enough tests
Silicon score74.072 to 76 with other tests
Price$0.10 / $0.50in / out, per million tokens
Each job on its own · scored in 5 of 6
Chat & writing67.6
Coding74.4
Math & reasoning74.0
Web research72.7
Understanding images74.0
Its tests
GPT-6 Luna took 21 of our tests, from 4 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
41.4%
26 of 32
11%
Arena SearchArena (LMArena)
1160 rating
17 of 17
11%
Arena VisionArena (LMArena)
1241 rating
29 of 29
11%
Arena AgentArena (LMArena)
+3.19 points
18 of 34
8%
Arena WebDevArena (LMArena)
1582 rating
27 of 37
7%
Arena TextArena (LMArena)
1443 rating
34 of 34
6%
FACTS SearchFACTS (Google DeepMind and Kaggle)
81.42%
8 of 13
5%
LiveBench ReasoningLiveBench
81.77%
31 of 34
5%
Furniture AssemblyEpoch AI
44.17%
12 of 28
5%
FACTS MultimodalFACTS (Google DeepMind and Kaggle)
34.25%
13 of 13
5%
LiveBench CodingLiveBench
78.95%
19 of 34
4%
LiveBench Agentic CodingLiveBench
51.21%
28 of 34
4%
LiveBench LanguageLiveBench
73.83%
32 of 34
3%
LiveBench Instruction FollowingLiveBench
55.93%
33 of 34
3%
FACTS GroundingFACTS (Google DeepMind and Kaggle)
81.37%
4 of 15
2%
FACTS ParametricFACTS (Google DeepMind and Kaggle)
56.08%
9 of 14
2%
FrontierMath Tiers 1–3Epoch AI
78.95%
15 of 34
2%
Mystery Game PuzzlesEpoch AI
7%
32 of 32
2%
LiveBench MathematicsLiveBench
89.12%
28 of 34
2%
GPQA DiamondEpoch AI
90.47%
27 of 32
1%
Chess PuzzlesEpoch AI
36%
29 of 35
1%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.
Moonshot AI (Kimi)Open model
Kimi K2.6
Place#33of 33 models with enough tests
Silicon score72.169.9 to 74.3 with other tests
PriceNo token priceno official price per token
Each job on its own · scored in 4 of 6
Chat & writing68.6
Coding68.2
Math & reasoning67.2
Understanding images77.8
Its tests
Kimi K2.6 took 17 of our tests, from 3 publishers. Its place is among the models we list that took the same test. The share says how much of the score each test decides.
Test
Result
Place
Share
SimpleQA VerifiedEpoch AI
34.9%
29 of 32
16%
Arena WebDevArena (LMArena)
1509 rating
36 of 37
10%
Arena TextArena (LMArena)
1461 rating
29 of 34
8%
LiveBench ReasoningLiveBench
79.38%
32 of 34
7%
Arena VisionArena (LMArena)
1265 rating
22 of 29
7%
Arena DocumentArena (LMArena)
1451 rating
16 of 17
7%
Furniture AssemblyEpoch AI
21.67%
26 of 28
7%
EBR-benchEpoch AI
2.38%
17 of 17
6%
LiveBench CodingLiveBench
78.57%
23 of 34
5%
LiveBench Agentic CodingLiveBench
46.92%
33 of 34
5%
LiveBench LanguageLiveBench
75.14%
29 of 34
4%
LiveBench Instruction FollowingLiveBench
64.36%
27 of 34
4%
FrontierMath Tiers 1–3Epoch AI
57.19%
31 of 34
3%
Mystery Game PuzzlesEpoch AI
18%
30 of 32
3%
LiveBench MathematicsLiveBench
84.28%
33 of 34
3%
GPQA DiamondEpoch AI
90.78%
22 of 32
2%
Chess PuzzlesEpoch AI
20%
35 of 35
2%
All its tests are put together in one score, and each of the 6 jobs counts the same, so the total is not a plain average of the job scores. The best model gets 100, and every 10 points less is one typical gap behind it.