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.
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.
Riemann-bench Leaderboard
1
GPT
5.6 Sol (Max reasoning)
74.4
%
1
GPT
6 Astra (Max reasoning)
72
%
1
Claude
Opus 5.5 (Adaptive/Max)
69.6
%
1
Claude
Opus 5 (Adaptive/Max)
68
%
1
GPT
5.6 Terra (Max reasoning)
66.4
%
1
Claude
Fable 5.1 (Adaptive/Max)
65.6
%
1
GPT
6 Sol (Max reasoning)
64.8
%
1
Claude
Fable 5 (Adaptive/Max)
60
%
1
GPT
6 Luna (Max reasoning)
59.2
%
1
GPT
5.5 (xHigh reasoning)
55.2
%
1
Gemini
3.8 Flash (High reasoning)
51.2
%
1
Claude
Opus 4.8 (Adaptive/Max)
47.2
%
1
GPT
5.4 (xHigh reasoning)
41.6
%
1
Gemini
3.7 Flash (High reasoning)
39.2
%
1
DeepSeek
V4 Pro (Max reasoning)
38.4
%
1
Grok
4.5 (High reasoning)
38.4
%
1
GPT
5.2 (xHigh reasoning)
37.6
%
1
Kimi
K3 (Max reasoning)
37.6
%
1
Gemini
3.5 Flash (High reasoning)
36.8
%
1
Gemini
3.1 Pro (High reasoning)
33.6
%
1
Claude
Opus 4.7 (Adaptive/Max)
32.8
%
1
Gemini
3.6 Flash (High reasoning)
30.4
%
1
Muse
Spark 1.3 (xHigh reasoning)
28
%
1
Claude
Opus 4.6 (Adaptive/Max)
27.2
%
1
DeepSeek
V4.1 Flash (Max reasoning)
24
%
1
Hy
Hy4 Preview (High reasoning)
24
%
1
DeepSeek
V4 Flash Vision (experimental) (High reasoning)
23.2
%
1
Muse
Spark 1.2 (xHigh reasoning)
23.2
%
1
Muse
Spark 1.1 (xHigh reasoning)
20.8
%
1
Qwen
3.7 Max (Thinking on)
15.2
%
1
Qwen
3.8 Max (xHigh reasoning)
15.2
%
1
Inkling
Inkling (xHigh reasoning)
15.2
%
1
Muse
Glimmer 30B (xHigh reasoning)
13.6
%
1
GLM
5.3 (Max reasoning)
12
%
1
Kimi
K2.5 (Thinking on)
12
%
1
Claude
Opus 4.5 (Adaptive/Max)
11.2
%
1
Gemini
3.5 Flash-Lite (High reasoning)
11.2
%
1
DeepSeek
V4 Flash (Max reasoning)
10.4
%
1
GLM
5.2 (Max reasoning)
10.4
%
1
GLM
5.3 Flash (Max reasoning)
8.8
%
1
DeepSeek
V3.2 (High reasoning)
8
%
1
Kimi
K2.6 (Thinking on)
7.2
%
1
DeepSeek
V4 Pro (preview) (xHigh reasoning)
5.6
%
1
MAI
Thinking 1
4.8
%
VIEW ALL BENCHMARKS
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.
Rank
Model
Score
41.7
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
41.7
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
33.9
%
GPT 6 Astra (Max)
OpenAI
31.1
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
28.9
%
GPT 6.1 Sol (Max)
OpenAI
20.6
%
Qwen 3.8 Max
Alibaba Cloud
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
24
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
23.9
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
21.5
%
GPT 6 Astra (Max)
OpenAI
21.3
%
Muse Wonder 1.0 (Max)
Meta
19.8
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
18.8
%
Gemini 4 Argon (High)
Google