RL Environments and the Hierarchy of Agentic Capabilities
Our RL environment run on 9 models revealed the core capabilities all agents need to master: tool use, planning, adaptability, groundedness, and common sense.
Professional Spreadsheet Reasoning
GDP.xlsx
Can frontier models master the spreadsheets that run the world? GDP.xlsx is a reasoning benchmark that takes real-world prompts and workbooks pulled directly from expert professional workflows, from financial models and budgets to schedules, inventories, and audit trails.
Leaderboard
No items found.
Sample tasks
Task Prompt
Write a tense dialogue between two siblings arguing over their late mother's estate. One is practical and business-minded, the other is sentimental and emotional. Show their different personalities through their speech patterns.
Final Response
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics failure reason
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Task Prompt
Write a tense dialogue between two siblings arguing over their late mother's estate. One is practical and business-minded, the other is sentimental and emotional. Show their different personalities through their speech patterns.
Final Response
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics failure reason
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Task Prompt
Write a tense dialogue between two siblings arguing over their late mother's estate. One is practical and business-minded, the other is sentimental and emotional. Show their different personalities through their speech patterns.
Final Response
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Rubrics failure reason
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
VIEW ALL BENCHMARKS
Agentic Healthcare
DAYJOB: Healthcare
DAYJOB: Healthcare evaluates long-horizon healthcare agents across clinical, operational, payer, pharmacy, and compliance workflows. Rather than telling models exactly what to do, it asks whether agents can figure out what needs to be done, and reason through it until the end.
Rank
Model
Score
24.7
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
11.6
%
GPT 6 Astra (Max)
OpenAI
9.6
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
8.4
%
Claude Opus 5 (Adaptive/Max)
Anthropic
8.4
%
Grok 4.7 (xHigh)
xAI
7.8
%
Muse Spark 1.3 (Max)
Meta
5.6
%
Claude Fable 5 (Adaptive/Max)
Anthropic
3.6
%
GPT 6 Sol (Max)
OpenAI
2.8
%
Grok 4.6 (xHigh)
xAI
2
%
Muse Spark 1.2 (xHigh)
Meta
2
%
Qwen 3.8 Max
Alibaba Cloud
1.6
%
GPT 5.6 Sol (Max)
OpenAI
1.6
%
GPT 5.6 Terra (xHigh)
OpenAI
1.2
%
Kimi K3 (Max)
Moonshot AI
0.8
%
Gemini 3.8 Flash (High)
Google
0.8
%
Hy4 Preview
Tencent
0.4
%
GLM 5.3 (Max)
Zhipu AI
0.4
%
GLM 5.3 Flash (Max)
Zhipu AI
0.4
%
GPT 5.6 Luna (Max)
OpenAI
0.4
%
DeepSeek V4 Pro (Max)
DeepSeek
0.4
%
Hy3 (High)
Tencent
0
%
Claude Sonnet 5 (Adaptive/Max)
Anthropic
0
%
Gemini 3.7 Flash (High)
Google
0
%
Inkling (Max)
Thinking Machines
0
%
DeepSeek V4 Flash (Max)
DeepSeek
0
%
Gemini 3.1 Pro
Google
0
%
Kimi K2.7 Code (Max)
Moonshot AI
0
%
Muse Glimmer 30B (xHigh)
Meta
0
%
Nemotron 3 Ultra
NVIDIA
0
%
Mistral Large 3
Mistral AI
0
%
GPT 6 Luna (Max)
OpenAI
Professional Graphical Reasoning
Chartography
Chartography is our benchmark for professional chart understanding. It tests whether frontier models can read the Kaplan-Meier curves, candlestick charts, contour maps, Sankey diagrams, Bode plots, and other specialized graphics that professionals use to make real decisions every day. It evaluates visual perception, domain-aware interpretation, and multi-step graphical reasoning.
Rank
Model
Score
71
%
GPT 6 Astra (Max)
OpenAI
66.3
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
53.6
%
GPT 6 Sol (Max)
OpenAI
46.2
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
45
%
GPT 5.6 Sol (Max)
OpenAI
40.9
%
Gemini 3.8 Flash (High)
Google
40.4
%
Gemini 3.7 Flash (High)
Google
35.9
%
Gemini 3.5 Flash
Google
34.8
%
Claude Fable 5 (Adaptive/Max)
Anthropic
34
%
GPT 5.6 Terra (Max)
OpenAI
34
%
Gemini 3.6 Flash
Google
32.1
%
Muse Spark 1.2 (xHigh)
Meta
31.4
%
GPT 5.6 Luna (Max)
OpenAI
31
%
GPT 5.5 (xHigh)
OpenAI
29.6
%
GPT 5.4 (xHigh)
OpenAI
29.1
%
Qwen 3.8 Max
Alibaba Cloud
29.1
%
GPT 6 Luna (Max)
OpenAI
27.6
%
Muse Spark 1.3 (xHigh)
Meta
27.3
%
Claude Opus 5 (Adaptive/Max)
Anthropic
26.6
%
Kimi K3 (Max)
Moonshot AI
26.1
%
Gemini 3.1 Pro
Google
24.4
%
Muse Spark 1.1 (xHigh)
Meta
20.5
%
Grok 4.6 (xHigh)
xAI
17.4
%
Qwen 3.8 Flash
Alibaba Cloud
17.2
%
Muse Glimmer 30B (xHigh)
Meta
17
%
Grok 4.5
xAI
16.6
%
Claude Sonnet 5 (Adaptive/Max)
Anthropic
16.5
%
Claude Opus 4.7 (Adaptive/Max)
Anthropic
16.3
%
GLM 5.3 Flash
Zhipu AI
15.9
%
Claude Opus 4.8 (Adaptive/Max)
Anthropic
15.9
%
Qwen 3.5 Plus
Alibaba Cloud
15.8
%
Qwen 3.7 Plus
Alibaba Cloud
14.7
%
Grok 4.7 (xHigh)
xAI
12.9
%
Grok 4.3 (High)
xAI
12.6
%
Kimi K2.5
Moonshot AI
12.2
%
Kimi K2.6
Moonshot AI
11.6
%
DeepSeek V4.1 Flash (Max)
DeepSeek
11.2
%
DeepSeek V4 Flash Vision (experimental) (Max)
DeepSeek
10.2
%
Gemini 3.5 Flash-Lite
Google
10
%
Inkling
Thinking Machines
9.1
%
Inkling Small (xHigh)
Thinking Machines
9
%
Mistral Large 3
Mistral AI
Long-Context Agentic Instruction Following
HANDBOOK.md Agents
Can an agent follow a 100-page company handbook inside an enterprise RL environment?
HANDBOOK.md is a benchmark for long-context agentic instruction following, modeled on how professionals follow corporate policy in their day-to-day work. Each task is a unique RL environment with internal tools and external MCP servers, across five enterprise domains.
Rank
Model
Score
38.8
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
36.2
%
Claude Fable 5 (Adaptive/Max)
Anthropic
32.3
%
Claude Opus 5 (Adaptive/Max)
Anthropic
27.3
%
Grok 4.6 (xHigh)
xAI
26.9
%
DeepSeek V4 Pro (Max)
DeepSeek
23.5
%
GPT 5.6 Sol (Max)
OpenAI
23.1
%
DeepSeek V4 Flash Vision (experimental) (Max)
DeepSeek
21.9
%
Claude Opus 4.8 (Adaptive/Max)
Anthropic
21.5
%
GPT 5.5 (xHigh)
OpenAI
16.5
%
Qwen 3.8 Max
Alibaba Cloud
15.8
%
Grok 4.5 (High)
xAI
15.4
%
Muse Spark 1.3 (xHigh)
Meta
15
%
GLM 5.3
Zhipu AI
13.5
%
Muse Spark 1.1 (xHigh)
Meta
13.1
%
Muse Spark 1.2 (xHigh)
Meta
11.9
%
Kimi K3 (Max)
Moonshot AI
11.2
%
Gemini 3.5 Flash (High)
Google
10.8
%
Gemini 3.8 Flash (High)
Google
10.8
%
Gemini 3.7 Flash (High)
Google
10.4
%
GLM 5.3 Flash
Zhipu AI
10.4
%
Claude Sonnet 4.6 (Adaptive/Max)
Anthropic
10
%
Gemini 3.1 Pro
Google
10
%
GLM 5.2 (xHigh)
Zhipu AI
9.2
%
DeepSeek V4 Pro (preview) (xHigh)
DeepSeek
8.5
%
Qwen 3.7 Max
Alibaba Cloud
7.7
%
Hy3 (High)
Tencent
7.3
%
DeepSeek V4 Flash (xHigh)
DeepSeek
7.3
%
DeepSeek V4 Flash (preview) (xHigh)
DeepSeek
6.9
%
Kimi K2.6
Moonshot AI
5
%
Gemini 3.6 Flash (High)
Google
3.5
%
Muse Glimmer 30B (xHigh)
Meta
3.1
%
Gemini 3.5 Flash-Lite (High)
Google
2.3
%
Inkling (xHigh)
Thinking Machines
1.9
%
Grok 4.3 (High)
xAI
1.5
%
Nemotron 3 Ultra
NVIDIA
1.2
%
Inkling Small (xHigh)
Thinking Machines
0
%
Nemotron Lightning 3.5 30B A3B
NVIDIA
Antidote / Everyday
Antidote: Everyday Edition
Some leaderboards reward the answers that convince you in two seconds. Antidote rewards the answer that leaves you better off a month later. It's our real-world AI leaderboard, graded by doctors, lawyers, and engineers who read every word, check every citation, and run every line of code.
Rank
Model
elo score (95% ci)
Claude Fable 5
1099
(
1085
-
1112
)
Anthropic
Gemini 3.1 Pro
1091
(
1081
-
1101
)
Google
Gemini 3.8 Flash
1087
(
1070
-
1104
)
Google
Gemini 3.6 Flash
1083
(
1069
-
1097
)
Google
Gemini 3.5 Flash
1079
(
1068
-
1090
)
Google
Kimi K3
1077
(
1063
-
1091
)
Moonshot AI
Gemini 3.7 Flash
1075
(
1059
-
1090
)
Google
GLM 5.3
1067
(
1051
-
1083
)
Zhipu AI
Claude Fable 5.1
1065
(
1048
-
1083
)
Anthropic
Claude Opus 5
1055
(
1041
-
1069
)
Anthropic
Qwen 3.7 Max
1051
(
1040
-
1061
)
Alibaba Cloud
Kimi K2.6
1045
(
1036
-
1054
)
Moonshot AI
Claude Opus 4.7
1042
(
1031
-
1053
)
Anthropic
Claude Opus 4.6
1042
(
1032
-
1052
)
Anthropic
GLM 5.2
1040
(
1027
-
1052
)
Zhipu AI
Claude Opus 4.8
1030
(
1018
-
1043
)
Anthropic
Muse Spark 1.2
1028
(
1011
-
1046
)
Meta
GPT 5.6 Sol
1027
(
1014
-
1041
)
OpenAI
Hy3
1022
(
1007
-
1037
)
Tencent
DeepSeek V4 Pro
1018
(
1003
-
1033
)
DeepSeek
DeepSeek V4 Flash Vision (experimental)
1015
(
999
-
1030
)
DeepSeek
Muse Spark 1.1
1014
(
998
-
1030
)
Meta
GPT 5.5
1013
(
1000
-
1025
)
OpenAI
Claude Sonnet 4.6
1012
(
1000
-
1024
)
Anthropic
DeepSeek V4 Pro (preview)
1007
(
995
-
1019
)
DeepSeek
Kimi K2.5
1006
(
995
-
1018
)
Moonshot AI
Grok 4.20 Beta
1001
(
989
-
1012
)
xAI
Qwen 3.5 Plus
999
(
987
-
1011
)
Alibaba Cloud
Hy4 Preview
992
(
974
-
1009
)
Tencent
Gemini 3.5 Flash-Lite
991
(
977
-
1005
)
Google
Inkling
991
(
977
-
1005
)
Thinking Machines
Qwen 3.8 Max
985
(
970
-
1000
)
Alibaba Cloud
Grok 4.5
973
(
960
-
986
)
xAI
DeepSeek V4 Flash (preview)
966
(
954
-
978
)
DeepSeek
Grok 4.6
966
(
950
-
981
)
xAI
DeepSeek V3.2
954
(
940
-
967
)
DeepSeek
Grok 4.3
950
(
939
-
960
)
xAI
Mistral Large 3
947
(
938
-
957
)
Mistral AI
Ernie 5.1
947
(
936
-
958
)
Baidu
Qwen 3.8 Flash
944
(
927
-
962
)
Alibaba Cloud
Claude Haiku 4.5
932
(
919
-
944
)
Anthropic
Muse Glimmer 30B
931
(
915
-
947
)
Meta
GPT 5.4 Mini
929
(
917
-
941
)
OpenAI
Gemma 3 12B
893
(
880
-
905
)
Google
Nemotron Lightning 3.5 30B A3B
888
(
871
-
904
)
NVIDIA
Ernie 4.5 300B
863
(
850
-
876
)
Baidu
Nova 2 Pro
819
(
809
-
830
)
Amazon
Creative, Business, and Everyday Writing
Hemingway-bench
Most AI writing benchmarks reward surface-level signals: elaborate metaphors and prose that looks impressive at a glance. Hemingway-bench rewards writing that is actually good.
Our leaderboard is judged by professional writers who evaluate creative writing, business writing, and everyday writing tasks for taste, originality, coherence, and emotional intelligence.
Rank
Model
elo score (95% ci)
Claude Fable 5
1110
(
1091
-
1130
)
Anthropic
Gemini 3.7 Flash
1105
(
1084
-
1126
)
Google
Claude Fable 5.1
1095
(
1071
-
1118
)
Anthropic
Gemini 3.8 Flash
1093
(
1069
-
1117
)
Google
Gemini 3.6 Flash
1075
(
1055
-
1095
)
Google
Gemini 3.1 Pro
1073
(
1057
-
1088
)
Google
Gemini 3.5 Flash
1071
(
1054
-
1088
)
Google
Kimi K3
1069
(
1049
-
1089
)
Moonshot AI
GPT 5.6 Sol
1057
(
1037
-
1076
)
OpenAI
Gemini 3 Pro
1056
(
1031
-
1081
)
Google
GLM 5.3
1053
(
1030
-
1076
)
Zhipu AI
Gemini 3 Flash
1052
(
1037
-
1066
)
Google
Claude Opus 5
1047
(
1026
-
1068
)
Anthropic
DeepSeek V4 Pro
1046
(
1025
-
1068
)
DeepSeek
Claude Opus 4.6
1045
(
1029
-
1061
)
Anthropic
Claude Opus 4.8 (Adaptive/Default)
1044
(
1026
-
1061
)
Anthropic
Claude Opus 4.7
1042
(
1026
-
1058
)
Anthropic
Claude Opus 4.8
1038
(
1021
-
1055
)
Anthropic
GLM 5.2
1036
(
1019
-
1053
)
Zhipu AI
GPT 5.5
1035
(
1019
-
1052
)
OpenAI
GLM 5.3 Flash
1034
(
1010
-
1057
)
Zhipu AI
Kimi K2.6
1021
(
1004
-
1038
)
Moonshot AI
Claude Opus 4.5
1016
(
995
-
1037
)
Anthropic
Hy4 Preview
1014
(
991
-
1037
)
Tencent
Muse Spark 1.2
1013
(
992
-
1034
)
Meta
Qwen 3.8 Max
1009
(
988
-
1030
)
Alibaba Cloud
Claude Sonnet 4.6
1001
(
985
-
1016
)
Anthropic
DeepSeek V4 Pro (preview)
999
(
982
-
1015
)
DeepSeek
GPT 5.2 Chat
994
(
978
-
1010
)
OpenAI
Gemini 3.5 Flash-Lite
990
(
970
-
1010
)
Google
Qwen 3.5 Plus
990
(
974
-
1005
)
Alibaba Cloud
Kimi K2.5
988
(
973
-
1003
)
Moonshot AI
Muse Spark 1.1
985
(
966
-
1005
)
Meta
GPT 5.4
985
(
969
-
1001
)
OpenAI
Grok 4.6
983
(
962
-
1004
)
xAI
DeepSeek V4 Flash Vision (experimental)
981
(
958
-
1003
)
DeepSeek
Grok 4.5
972
(
954
-
991
)
xAI
Hy3
971
(
949
-
994
)
Tencent
DeepSeek V4 Flash (preview)
969
(
953
-
986
)
DeepSeek
GPT 5.2
954
(
929
-
980
)
OpenAI
Qwen 3 Max
948
(
920
-
975
)
Alibaba Cloud
Qwen 3.8 Flash
929
(
905
-
952
)
Alibaba Cloud
Grok 4.1 Fast Reasoning
916
(
901
-
931
)
xAI
Muse Glimmer 30B
913
(
892
-
935
)
Meta
Nemotron Lightning 3.5 30B A3B
888
(
866
-
911
)
NVIDIA
Kimi K2 Instruct
883
(
856
-
910
)
Moonshot AI
Llama 4 Maverick
803
(
781
-
824
)
Meta
Nova 2 Pro
749
(
718
-
779
)
Amazon
Agentic Finance
DAYJOB: Finance
DAYJOB: Finance measures long-horizon finance agents across corporate finance, banking, credit, investing, and real assets. It's built around the messy reality of knowledge work: realistic requests, complex environments, and end-to-end professional judgment.
Rank
Model
Score
23.9
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
21.5
%
GPT 6 Astra (Max)
OpenAI
19.8
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
14.8
%
Muse Spark 1.3 (Max)
Meta
14.5
%
Grok 4.7 (xHigh)
xAI
12.8
%
Grok 4.6 (xHigh)
xAI
11.3
%
Claude Opus 5 (Adaptive/Max)
Anthropic
9.3
%
GPT 6 Sol (Max)
OpenAI
9
%
Claude Fable 5 (Adaptive/Max)
Anthropic
7.3
%
Hy4 Preview
Tencent
6
%
GLM 5.3 (Max)
Zhipu AI
5.5
%
GPT 5.6 Sol (Max)
OpenAI
3.8
%
Kimi K3 (Max)
Moonshot AI
3
%
Qwen 3.8 Max
Alibaba Cloud
2.8
%
Gemini 3.7 Flash (High)
Google
2.8
%
Gemini 3.8 Flash (High)
Google
2.3
%
GLM 5.3 Flash (Max)
Zhipu AI
2.3
%
Muse Spark 1.2 (xHigh)
Meta
2.3
%
Claude Sonnet 5 (Adaptive/Max)
Anthropic
2.3
%
GPT 6 Luna (Max)
OpenAI
1.8
%
GPT 5.6 Luna (Max)
OpenAI
1.8
%
GPT 5.6 Terra (xHigh)
OpenAI
1
%
DeepSeek V4 Pro (Max)
DeepSeek
0.3
%
Hy3 (High)
Tencent
0
%
DeepSeek V4 Flash (Max)
DeepSeek
0
%
Gemini 3.1 Pro
Google
0
%
Inkling (Max)
Thinking Machines
0
%
Kimi K2.7 Code (Max)
Moonshot AI
0
%
Mistral Large 3
Mistral AI
0
%
Muse Glimmer 30B (xHigh)
Meta
0
%
Nemotron 3 Ultra
NVIDIA
Enterprise Instruction Following
ComplexConstraints
A benchmark for professional instruction following, where constraints depend on each other, fire conditionally, and must be inferred from context.
Rank
Model
Score
GPT 6 Astra (Max)
57.7
%
OpenAI
Muse Spark 1.3 (xHigh)
51.9
%
Meta
GPT 6 Sol (Max)
50.8
%
OpenAI
GPT 5.6 Sol (Max)
50.5
%
OpenAI
Claude Opus 5.5 (Adaptive/Max)
49.9
%
Anthropic
DeepSeek V4.1 Flash (Max)
49.7
%
DeepSeek
GPT 5.5 (xHigh)
49.5
%
OpenAI
Grok 4.7 (xHigh)
48.9
%
xAI
Gemini 3.8 Flash (High)
48.4
%
Google
GPT 5.4 (xHigh)
46
%
OpenAI
Hy4 Preview
46
%
Tencent
Qwen 3.8 Max
45.5
%
Alibaba Cloud
Claude Fable 5.1 (Adaptive/Max)
45.1
%
Anthropic
Gemini 3.1 Pro
43.7
%
Google
Qwen 3.8 Flash
43.3
%
Alibaba Cloud
Gemini 3.7 Flash (High)
42.4
%
Google
DeepSeek V4 Pro (Max)
42
%
DeepSeek
Gemini 3.6 Flash (High)
40
%
Google
Muse Spark 1.2 (xHigh)
39.9
%
Meta
DeepSeek V4 Flash Vision (experimental) (Max)
39.9
%
DeepSeek
Claude Fable 5 (Max)
38.1
%
Anthropic
Muse Spark 1.1 (xHigh)
38.1
%
Meta
Kimi K3
37.9
%
Moonshot AI
Claude Opus 5 (Adaptive/Max)
37.3
%
Anthropic
Gemini 3.5 Flash
37.1
%
Google
Grok 4.6 (xHigh)
36.5
%
xAI
Claude Opus 4.6 (Adaptive/High)
36.3
%
Anthropic
Grok 4.5
35.9
%
xAI
Claude Opus 4.8 (Max)
35.6
%
Anthropic
Hy3 (High)
34.3
%
Tencent
Inkling
34.1
%
Thinking Machines
Claude Sonnet 4.6 (Adaptive/High)
34
%
Anthropic
Qwen 3.7 Max
33.5
%
Alibaba Cloud
Claude Opus 4.7
31.6
%
Anthropic
GLM 5.3
30.5
%
Zhipu AI
Claude Opus 4.7 (Adaptive/High)
30.3
%
Anthropic
Kimi K2.6
30.1
%
Moonshot AI
GLM 5.2 (Max)
29.3
%
Zhipu AI
DeepSeek V4 Pro (preview)
28
%
DeepSeek
GLM 5.3 Flash
27.6
%
Zhipu AI
GPT 6 Luna
27.3
%
OpenAI
Muse Glimmer 30B (xHigh)
23.9
%
Meta
Gemini 3.5 Flash-Lite (High)
23.2
%
Google
DeepSeek V4 Flash (preview)
20.9
%
DeepSeek
Kimi K2.5
18.2
%
Moonshot AI
Qwen 3.5 Plus
17.8
%
Alibaba Cloud
Nemotron 3 Ultra
17.8
%
NVIDIA
Claude Opus 4.6
16.9
%
Anthropic
Grok 4.20 Beta
16.4
%
xAI
Ernie 5.1
14.7
%
Baidu
Claude Sonnet 4.6
10.7
%
Anthropic
Nemotron Lightning 3.5 30B A3B
7.1
%
NVIDIA
DeepSeek V3.2
2.2
%
DeepSeek
Mistral Large 3
0.4
%
Mistral AI
Ernie 4.5 300B
0
%
Baidu
Nova 2 Pro
0
%
Amazon
Professional Multimodal Reasoning
GDP.pdf
Can frontier models master the documents that run the world? GDP.pdf is a multimodal and reasoning benchmark that takes real-world prompts and PDFs pulled directly from expert professional workflows.
Rank
Model
Score
34.2
%
GPT 6 Astra (Max)
OpenAI
30.7
%
GPT 5.6 Sol (Max)
OpenAI
30.6
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
29.8
%
Claude Fable 5 (Adaptive/Max)
Anthropic
27.6
%
Muse Spark 1.3 (xHigh)
Meta
27.6
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
26.4
%
GPT 6 Sol (Max)
OpenAI
26
%
GPT 5.5 (xHigh)
OpenAI
24.7
%
GPT 5.6 Terra
OpenAI
24
%
Claude Opus 5 (Adaptive/Max)
Anthropic
24
%
Claude Opus 4.8 (Adaptive/Max)
Anthropic
23.8
%
Gemini 3.7 Flash (High)
Google
23.2
%
Gemini 3.8 Flash (High)
Google
23.2
%
Qwen 3.8 Max
Alibaba Cloud
23
%
GPT 6 Luna (Max)
OpenAI
22.8
%
Grok 4.7 (xHigh)
xAI
22.7
%
GPT 5.6 Luna
OpenAI
21
%
Claude Opus 4.7 (Adaptive/Max)
Anthropic
19.8
%
DeepSeek V4.1 Flash (Max)
DeepSeek
19
%
Kimi K3 (Max)
Moonshot AI
18
%
Claude Sonnet 4.6 (Adaptive/Max)
Anthropic
17.2
%
Grok 4.6 (xHigh)
xAI
17
%
Gemini 3.1 Pro
Google
16.6
%
Qwen 3.8 Flash
Alibaba Cloud
16.2
%
DeepSeek V4 Flash Vision (experimental) (Max)
DeepSeek
16
%
Muse Spark 1.2 (xHigh)
Meta
15
%
Muse Spark 1.1
Meta
14
%
GLM 5.3 Flash
Zhipu AI
14
%
Grok 4.5 (High)
xAI
14
%
Gemini 3.6 Flash (High)
Google
14
%
Gemini 3.5 Flash
Google
12
%
Kimi K2.6
Moonshot AI
11.8
%
Muse Glimmer 30B (xHigh)
Meta
10
%
Gemini 3.5 Flash-Lite (High)
Google
10
%
Gemini 3 Flash
Google
8
%
Grok 4.3 (High)
xAI
2
%
Nova 2 Pro
Amazon
2
%
Nemotron 3 Nano Omni
NVIDIA
2
%
Mistral Large 3
Mistral AI
Enterprise Agents in Realistic RL Environments
EnterpriseBench: CoreCraft Agents
Stop testing models in tiny, self-contained environments. We built CoreCraft, a large-scale startup world, and deployed AI agents to solve real tasks. Our goal: to move agents beyond the cleanliness of the lab and into the chaos of enterprise reality.
Rank
Model
Score
77.4
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
74.4
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
70.3
%
Claude Fable 5 (Adaptive/Max)
Anthropic
Frontier Research Mathematics
Riemann-bench
We evaluate AI models on advanced mathematical problems requiring deep reasoning and novel synthesis. Our benchmark features cutting-edge problems sourced from leading mathematicians, including Ivy League professors, PhD IMO medalists, and graduate students at the top of their field.
Rank
Model
Score
74.4
%
GPT-5.6 Sol (Max)
OpenAI
72
%
GPT 6 Astra (Max)
OpenAI
69.6
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
68
%
Claude Opus 5 (Adaptive/Max)
Anthropic
66.4
%
GPT 5.6 Terra (Max)
OpenAI
65.6
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
64.8
%
GPT 6 Sol (Max)
OpenAI
60
%
Claude Fable 5 (Adaptive/Max)
Anthropic
59.2
%
GPT 6 Luna (Max)
OpenAI
55.2
%
GPT 5.5 (xHigh)
OpenAI
51.2
%
Gemini 3.8 Flash (High)
Google
47.2
%
Claude Opus 4.8 (Adaptive/Max)
Anthropic
41.6
%
GPT 5.4 (xHigh)
OpenAI
39.2
%
Gemini 3.7 Flash (High)
Google
38.4
%
Grok 4.5 (High)
xAI
38.4
%
DeepSeek V4 Pro (Max)
DeepSeek
37.6
%
Kimi K3
Moonshot AI
37.6
%
GPT-5.2 (xHigh)
OpenAI
36.8
%
Gemini 3.5 Flash (High)
Google
33.6
%
Gemini 3.1 Pro
Google
32.8
%
Claude Opus 4.7 (Adaptive/Max)
Anthropic
30.4
%
Gemini 3.6 Flash (High)
Google
28
%
Muse Spark 1.3 (xHigh)
Meta
27.2
%
Claude Opus 4.6 (Adaptive/Max)
Anthropic
24
%
Hy4 Preview
Tencent
24
%
DeepSeek V4.1 Flash (Max)
DeepSeek
23.2
%
Muse Spark 1.2 (xHigh)
Meta
23.2
%
DeepSeek V4 Flash Vision (experimental) (High)
DeepSeek
20.8
%
Muse Spark 1.1 (xHigh)
Meta
15.2
%
Inkling (xHigh)
Thinking Machines
15.2
%
Qwen 3.8 Max
Alibaba Cloud
15.2
%
Qwen 3.7 Max
Alibaba Cloud
13.6
%
Muse Glimmer 30B (xHigh)
Meta
12
%
Kimi K2.5
Moonshot AI
12
%
GLM 5.3 (Max)
Zhipu AI
11.2
%
Gemini 3.5 Flash-Lite (High)
Google
11.2
%
Claude Opus 4.5 (Adaptive/Max)
Anthropic
10.4
%
GLM 5.2 (Max)
Zhipu AI
10.4
%
DeepSeek V4 Flash (Max)
DeepSeek
8.8
%
GLM 5.3 Flash (Max)
Zhipu AI
8
%
DeepSeek V3.2 (Thinking)
DeepSeek
7.2
%
Kimi K2.6
Moonshot AI
5.6
%
DeepSeek V4 Pro (preview) (xHigh)
DeepSeek
4.8
%
MAI Thinking 1
Microsoft