New Frontier data and RL environments, off the shelf

Blog
benchmarks
Off-the-Shelf
Training
experts
Products
Enterprise
Research
Careers
Contact
Claude Sonet 4.5
Accuracy of historical events, dates, and figures
EVALUATION DIMENSIONS
Adherence
91.8
Utility
87.2
Harmlessness
95.3
Reasoning
88.4
Robustness
84.2
Multimodal
74.5
Preference
89.5
Domain
84.9
Factuality
92.1
Structure
86.7
Dialogue
87.9
Patterns
81.6
Surge benchmark
85.7/100
About this score
The Human Judgment Score (HJS) is calculated as the average of twelve evaluation dimensions. Each dimension is scored 0-100 by expert human annotators using standardized protocols, providing a comprehensive assessment of model capabilities.

These dimensions measure critical aspects including accuracy, reasoning, safety, utility, and conversational quality. Higher scores indicate stronger alignment with human expectations and superior real-world performance.

Benchmarks

Greatness isn't accidental. How we measure it shouldn't be either. If we want AGI that builds billion-dollar enterprises and globe-spanning infrastructure, we need benchmarks that test for intelligence and sophistication.

This is our ranking of models, measured by their capacity for rigorous reasoning and real-world mastery.
View by :
Agentic Coding

sudo L7

Coding agents are getting very good at being L3s. But do they have the judgment to become L7s? Across 60 tasks, agents must navigate unfamiliar systems, choose architectures, spot hidden risks, and know when technically correct code is still the wrong thing to ship.

Leaderboard
Blog
Rank
Model
Score
Claude Sonnet 5.5 (Adaptive/Max)
45
%
Claude Sonnet 5.5 (Adaptive/Max)
Claude Opus 5.5 (Adaptive/Max)
45
%
Claude Opus 5.5 (Adaptive/Max)
GPT 6 Astra (Max)
32.2
%
GPT 6 Astra (Max)
GPT 6.1 Sol (Max)
31.7
%
GPT 6.1 Sol (Max)
Claude Fable 5.1 (Adaptive/Max)
27.2
%
Claude Fable 5.1 (Adaptive/Max)
Gemini 4 Argon (High)
19.4
%
Gemini 4 Argon (High)
Professional Spreadsheet Reasoning

GDP.xlsx

A spreadsheet might hold a company’s financial model, a laboratory’s results, or an engineer’s calculations. GDP.xlsx puts models to work inside these files, testing whether they can follow the evidence across a workbook and produce answers a professional could rely on.

Leaderboard
Blog
Hugging Face
Github
Rank
Model
Score
Gemini 4 Argon (High)
38.3
%
Gemini 4 Argon (High)
Claude Opus 5.5 (Adaptive/Max)
30.3
%
Claude Opus 5.5 (Adaptive/Max)
Claude Sonnet 5.5
29.1
%
Claude Sonnet 5.5
Claude Fable 5.1 (Adaptive/Max)
23.1
%
Claude Fable 5.1 (Adaptive/Max)
GPT 6 Astra (Max)
22.9
%
GPT 6 Astra (Max)
Muse Spark 1.3 (Max)
22.6
%
Muse Spark 1.3 (Max)
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.

Leaderboard
Blog
Research
Hugging Face
Github
Rank
Model
Score
Claude Opus 5.5 (Adaptive/Max)
24.7
%
Claude Opus 5.5 (Adaptive/Max)
GPT 6 Astra (Max)
11.6
%
GPT 6 Astra (Max)
Claude Fable 5.1 (Adaptive/Max)
9.6
%
Claude Fable 5.1 (Adaptive/Max)
Gemini 4 Argon (High)
9.6
%
Gemini 4 Argon (High)
Claude Opus 5 (Adaptive/Max)
8.4
%
Claude Opus 5 (Adaptive/Max)
Grok 4.7 (xHigh)
8.4
%
Grok 4.7 (xHigh)
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.

Leaderboard
Blog
Research
Github
Rank
Model
Score
Gemini 4 Argon (High)
71.6
%
Gemini 4 Argon (High)
GPT 6 Astra (Max)
71
%
GPT 6 Astra (Max)
Claude Opus 5.5 (Adaptive/Max)
66.3
%
Claude Opus 5.5 (Adaptive/Max)
GPT 6 Sol (Max)
53.6
%
GPT 6 Sol (Max)
Claude Fable 5.1 (Adaptive/Max)
46.2
%
Claude Fable 5.1 (Adaptive/Max)
GPT 5.6 Sol (Max)
45
%
GPT 5.6 Sol (Max)
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.

Leaderboard
Blog
Research
Github
Rank
Model
Score
Claude Fable 5.1 (Adaptive/Max)
38.8
%
Claude Fable 5.1 (Adaptive/Max)
Claude Fable 5 (Adaptive/Max)
36.2
%
Claude Fable 5 (Adaptive/Max)
Claude Opus 5 (Adaptive/Max)
32.3
%
Claude Opus 5 (Adaptive/Max)
Grok 4.6 (xHigh)
27.3
%
Grok 4.6 (xHigh)
DeepSeek V4 Pro (Max)
26.9
%
DeepSeek V4 Pro (Max)
GPT 5.6 Sol (Max)
23.5
%
GPT 5.6 Sol (Max)
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.

Leaderboard
Blog
Rank
Model
elo score (95% ci)
Claude Fable 5
1099
(
1085
-
1112
)
Gemini 3.1 Pro
1091
(
1081
-
1101
)
Gemini 3.8 Flash
1087
(
1070
-
1104
)
Gemini 3.6 Flash
1083
(
1069
-
1097
)
Gemini 3.5 Flash
1079
(
1068
-
1090
)
Kimi K3
1077
(
1063
-
1091
)
Gemini 3.7 Flash
1075
(
1059
-
1090
)
GLM 5.3
1067
(
1051
-
1083
)
Claude Fable 5.1
1065
(
1048
-
1083
)
Claude Opus 5
1055
(
1041
-
1069
)
Qwen 3.7 Max
1051
(
1040
-
1061
)
Kimi K2.6
1045
(
1036
-
1054
)
Claude Opus 4.7
1042
(
1031
-
1053
)
Claude Opus 4.6
1042
(
1032
-
1052
)
GLM 5.2
1040
(
1027
-
1052
)
Claude Opus 4.8
1030
(
1018
-
1043
)
Muse Spark 1.2
1028
(
1011
-
1046
)
GPT 5.6 Sol
1027
(
1014
-
1041
)
Hy3
1022
(
1007
-
1037
)
DeepSeek V4 Pro
1018
(
1003
-
1033
)
DeepSeek V4 Flash Vision (experimental)
1015
(
999
-
1030
)
Muse Spark 1.1
1014
(
998
-
1030
)
GPT 5.5
1013
(
1000
-
1025
)
Claude Sonnet 4.6
1012
(
1000
-
1024
)
DeepSeek V4 Pro (preview)
1007
(
995
-
1019
)
Kimi K2.5
1006
(
995
-
1018
)
Grok 4.20 Beta
1001
(
989
-
1012
)
Qwen 3.5 Plus
999
(
987
-
1011
)
Hy4 Preview
992
(
974
-
1009
)
Gemini 3.5 Flash-Lite
991
(
977
-
1005
)
Inkling
991
(
977
-
1005
)
Qwen 3.8 Max
985
(
970
-
1000
)
Grok 4.5
973
(
960
-
986
)
DeepSeek V4 Flash (preview)
966
(
954
-
978
)
Grok 4.6
966
(
950
-
981
)
DeepSeek V3.2
954
(
940
-
967
)
Grok 4.3
950
(
939
-
960
)
Mistral Large 3
947
(
938
-
957
)
Ernie 5.1
947
(
936
-
958
)
Qwen 3.8 Flash
944
(
927
-
962
)
Claude Haiku 4.5
932
(
919
-
944
)
Muse Glimmer 30B
931
(
915
-
947
)
GPT 5.4 Mini
929
(
917
-
941
)
Gemma 3 12B
893
(
880
-
905
)
Nemotron Lightning 3.5 30B A3B
888
(
871
-
904
)
Ernie 4.5 300B
863
(
850
-
876
)
Nova 2 Pro
819
(
809
-
830
)
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.

Leaderboard
Blog
Research
Hugging Face
Github
Rank
Model
Score
Claude Opus 5.5 (Adaptive/Max)
23.9
%
Claude Opus 5.5 (Adaptive/Max)
GPT 6 Astra (Max)
21.5
%
GPT 6 Astra (Max)
Gemini 4 Argon (High)
20.3
%
Gemini 4 Argon (High)
Claude Fable 5.1 (Adaptive/Max)
19.8
%
Claude Fable 5.1 (Adaptive/Max)
Muse Spark 1.3 (Max)
14.8
%
Muse Spark 1.3 (Max)
Grok 4.7 (xHigh)
14.5
%
Grok 4.7 (xHigh)
Enterprise Instruction Following

ComplexConstraints

A benchmark for professional instruction following, where constraints depend on each other, fire conditionally, and must be inferred from context.

Leaderboard
Blog
Research
Github
Rank
Model
Score
GPT 6 Astra (Max)
57.7
%
Gemini 4 Argon (High)
53.3
%
Muse Spark 1.3 (xHigh)
51.9
%
GPT 6 Sol (Max)
50.8
%
GPT 5.6 Sol (Max)
50.5
%
Claude Opus 5.5 (Adaptive/Max)
49.9
%
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.

Leaderboard
Blog
Research
Github
Rank
Model
Score
GPT 6.1 Sol (Max)
35
%
GPT 6.1 Sol (Max)
GPT 6 Astra (Max)
34.2
%
GPT 6 Astra (Max)
Muse Wonder Full (Max)
33.2
%
Muse Wonder Full (Max)
Claude Sonnet 5.5 (Adaptive/Max)
31.4
%
Claude Sonnet 5.5 (Adaptive/Max)
Muse Wonder (Max)
30.8
%
Muse Wonder (Max)
GPT 5.6 Sol (Max)
30.7
%
GPT 5.6 Sol (Max)
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.

Leaderboard
Blog
Research
Rank
Model
Score
Claude Fable 5.1 (Adaptive/Max)
77.4
%
Claude Fable 5.1 (Adaptive/Max)
Claude Opus 5.5 (Adaptive/Max)
74.4
%
Claude Opus 5.5 (Adaptive/Max)
Claude Fable 5 (Adaptive/Max)
70.3
%
Claude Fable 5 (Adaptive/Max)
Claude Opus 5 (Adaptive/Max)
68.7
%
Claude Opus 5 (Adaptive/Max)
Grok 4.7 (xHigh)
68.2
%
Grok 4.7 (xHigh)
Grok 4.6 (xHigh)
65.6
%
Grok 4.6 (xHigh)
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.

Leaderboard
Blog
Research
Rank
Model
Score
GPT-5.6 Sol (Max)
74.4
%
GPT-5.6 Sol (Max)
GPT 6 Astra (Max)
72
%
GPT 6 Astra (Max)
Claude Opus 5.5 (Adaptive/Max)
69.6
%
Claude Opus 5.5 (Adaptive/Max)
Claude Opus 5 (Adaptive/Max)
68
%
Claude Opus 5 (Adaptive/Max)
GPT 5.6 Terra (Max)
66.4
%
GPT 5.6 Terra (Max)
Claude Fable 5.1 (Adaptive/Max)
65.6
%
Claude Fable 5.1 (Adaptive/Max)
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.

Leaderboard
Blog
Rank
Model
elo score (95% ci)
Claude Fable 5
1110
(
1091
-
1130
)
Gemini 3.7 Flash
1105
(
1084
-
1126
)
Claude Fable 5.1
1095
(
1071
-
1118
)
Gemini 3.8 Flash
1093
(
1069
-
1117
)
Gemini 3.6 Flash
1075
(
1055
-
1095
)
Gemini 3.1 Pro
1073
(
1057
-
1088
)
Gemini 3.5 Flash
1071
(
1054
-
1088
)
Kimi K3
1069
(
1049
-
1089
)
GPT 5.6 Sol
1057
(
1037
-
1076
)
Gemini 3 Pro
1056
(
1031
-
1081
)
GLM 5.3
1053
(
1030
-
1076
)
Gemini 3 Flash
1052
(
1037
-
1066
)
Claude Opus 5
1047
(
1026
-
1068
)
DeepSeek V4 Pro
1046
(
1025
-
1068
)
Claude Opus 4.6
1045
(
1029
-
1061
)
Claude Opus 4.8 (Adaptive/Default)
1044
(
1026
-
1061
)
Claude Opus 4.7
1042
(
1026
-
1058
)
Claude Opus 4.8
1038
(
1021
-
1055
)
GLM 5.2
1036
(
1019
-
1053
)
GPT 5.5
1035
(
1019
-
1052
)
GLM 5.3 Flash
1034
(
1010
-
1057
)
Kimi K2.6
1021
(
1004
-
1038
)
Claude Opus 4.5
1016
(
995
-
1037
)
Hy4 Preview
1014
(
991
-
1037
)
Muse Spark 1.2
1013
(
992
-
1034
)
Qwen 3.8 Max
1009
(
988
-
1030
)
Claude Sonnet 4.6
1001
(
985
-
1016
)
DeepSeek V4 Pro (preview)
999
(
982
-
1015
)
GPT 5.2 Chat
994
(
978
-
1010
)
Gemini 3.5 Flash-Lite
990
(
970
-
1010
)
Qwen 3.5 Plus
990
(
974
-
1005
)
Kimi K2.5
988
(
973
-
1003
)
Muse Spark 1.1
985
(
966
-
1005
)
GPT 5.4
985
(
969
-
1001
)
Grok 4.6
983
(
962
-
1004
)
DeepSeek V4 Flash Vision (experimental)
981
(
958
-
1003
)
Grok 4.5
972
(
954
-
991
)
Hy3
971
(
949
-
994
)
DeepSeek V4 Flash (preview)
969
(
953
-
986
)
GPT 5.2
954
(
929
-
980
)
Qwen 3 Max
948
(
920
-
975
)
Qwen 3.8 Flash
929
(
905
-
952
)
Grok 4.1 Fast Reasoning
916
(
901
-
931
)
Muse Glimmer 30B
913
(
892
-
935
)
Nemotron Lightning 3.5 30B A3B
888
(
866
-
911
)
Kimi K2 Instruct
883
(
856
-
910
)
Llama 4 Maverick
803
(
781
-
824
)
Nova 2 Pro
749
(
718
-
779
)

Stay Posted

© 2026 Surge AI
X
Linkedin
Cookie Settings
Cookie Policy
Privacy Policy
Terms of Service