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.
ComplexConstraints Leaderboard
1
GPT
6 Astra (Max reasoning)
57.7
%
1
Gemini
4 Argon (High reasoning)
53.3
%
1
Muse
Spark 1.3 (xHigh reasoning)
51.9
%
1
GPT
6 Sol (Max reasoning)
50.8
%
1
GPT
5.6 Sol (Max reasoning)
50.5
%
1
Claude
Opus 5.5 (Adaptive/Max)
49.9
%
1
DeepSeek
V4.1 Flash (Max reasoning)
49.7
%
1
GPT
5.5 (xHigh reasoning)
49.5
%
1
Grok
4.7 (xHigh reasoning)
48.9
%
1
Gemini
3.8 Flash (High reasoning)
48.4
%
1
Hy
Hy4 Preview (High reasoning)
46
%
1
GPT
5.4 (xHigh reasoning)
46
%
1
Qwen
3.8 Max (xHigh reasoning)
45.5
%
1
Claude
Fable 5.1 (Adaptive/Max)
45.1
%
1
Gemini
3.1 Pro (High reasoning)
43.7
%
1
Qwen
3.8 Flash (xHigh reasoning)
43.3
%
1
Gemini
3.7 Flash (High reasoning)
42.4
%
1
DeepSeek
V4 Pro (Max reasoning)
42
%
1
Gemini
3.6 Flash (High reasoning)
40
%
1
DeepSeek
V4 Flash Vision (experimental) (Max reasoning)
39.9
%
1
Muse
Spark 1.2 (xHigh reasoning)
39.9
%
1
Muse
Spark 1.1 (xHigh reasoning)
38.1
%
1
Claude
Fable 5 (Max reasoning)
38.1
%
1
Kimi
K3 (Max reasoning)
37.9
%
1
Claude
Opus 5 (Adaptive/Max)
37.3
%
1
Grok
4.6 (xHigh reasoning)
36.5
%
1
Grok
4.5 (High reasoning)
35.9
%
1
Claude
Opus 4.8 (Max reasoning)
35.6
%
1
Hy
Hy3 (High reasoning)
34.3
%
1
Qwen
3.7 Max (Thinking on)
33.5
%
1
GLM
5.3 (Max reasoning)
30.5
%
1
Kimi
K2.6 (Thinking on)
30.1
%
1
GLM
5.2 (Max reasoning)
29.3
%
1
GLM
5.3 Flash (Max reasoning)
27.6
%
1
Muse
Glimmer 30B (xHigh reasoning)
23.9
%
1
Gemini
3.5 Flash-Lite (High reasoning)
23.2
%
1
Kimi
K2.5 (Thinking on)
18.2
%
1
Nemotron
3 Ultra
17.8
%
1
Qwen
3.5 Plus (Thinking on)
17.8
%
1
Grok
4.20 Beta
16.4
%
1
Ernie
5.1
14.7
%
1
Nemotron
Lightning 3.5 30B A3B (Thinking on)
7.1
%
1
Mistral
Large 3
0.4
%
1
Ernie
4.5 300B
0
%
VIEW ALL BENCHMARKS
Agentic Tax
DAYJOB: Tax
DAYJOB: Tax puts agents to work on 50 professional tax assignments spanning 1,789 files: individual returns, partnerships, estates, international taxation, and more. Agents must investigate client records, identify issues the client may have overlooked, and produce reliable tax work. The strongest model, Claude Sonnet 5.5, passes 10.8% of attempts at a 95% rubric threshold.
Rank
Model
Score
10.8
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
1
Claude Sonnet 5.5 (Adaptive/Max)
Sonnet 5.5 (Adaptive/Max)
10.8
5728d5ac-1252-4c29-9482-b9810698ddd6
9.6
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
2
Claude Fable 5.1 (Adaptive/Max)
Fable 5.1 (Adaptive/Max)
9.6
306e34bd-38ee-473e-b8ee-5ab34c61b86a
9.2
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
3
Claude Opus 5.5 (Adaptive/Max)
Opus 5.5 (Adaptive/Max)
9.2
f6ac0e88-3b86-4686-bc94-180b522259a7
6
%
Muse Wonder 1.0 (Max)
Meta
4
Muse Wonder 1.0 (Max)
Wonder 1.0 (Max reasoning)
6
a38a3c03-14ff-4fa1-ae22-e23b334c9102
5
%
Grok 4.7 (xHigh)
xAI
5
Grok 4.7 (xHigh)
4.7 (xHigh reasoning)
5
304e38c8-6587-436a-9349-aff803f5d2e7
4.8
%
GPT 6 Astra (Max)
OpenAI
6
GPT 6 Astra (Max)
6 Astra (Max reasoning)
4.8
f06fbd6c-999b-4ec9-a438-a6bae115d8b5
4
%
GPT 6.1 Sol (Max)
OpenAI
7
GPT 6.1 Sol (Max)
6.1 Sol (Max reasoning)
4
b0ca433b-1d4e-4532-968e-a7dca80ebb4a
3.6
%
Muse Spark 1.3 (Max)
Meta
8
Muse Spark 1.3 (Max)
Spark 1.3 (Max reasoning)
3.6
3c4d9355-ed31-47d3-889e-ef41c1b876e8
3.2
%
Gemini 4 Argon (High)
Google
9
Gemini 4 Argon (High)
4 Argon (High reasoning)
3.2
4af05e84-b2a4-438a-a8f5-6dc20704ba93
1.6
%
Hy4 Preview
Tencent
10
Hy4 Preview
Hy4 Preview (High reasoning)
1.6
8242f1b1-b2e6-43b4-bb9d-26e3ec9cc389
1.2
%
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek
11
DeepSeek V4.1 Flash 1017 (Max)
V4.1 Flash 1017 (Max reasoning)
1.2
68d35d26-f7cc-472c-907b-fccc62cae530
0.8
%
Gemini 3.8 Flash (High)
Google
12
Gemini 3.8 Flash (High)
3.8 Flash (High reasoning)
0.8
0dec1c04-a7f9-4674-8f75-ba3fedeff52d
0.4
%
GLM 5.3 (Max)
Zhipu AI
13
GLM 5.3 (Max)
5.3 (Max reasoning)
0.4
e7e6b0c4-3820-4b77-8e3f-4e86e795f999
0
%
Kimi K3 (Max)
Moonshot AI
14
Kimi K3 (Max)
K3 (Max reasoning)
0
11f34137-f8fb-4918-890b-7db0d3b65b05
0
%
Mistral Large 4 (High)
Mistral AI
14
Mistral Large 4 (High)
Large 4 (High reasoning)
0
b9e22405-d470-4b17-8c48-36cc6157a286
0
%
Qwen 3.8 Max
Alibaba Cloud
14
Qwen 3.8 Max
3.8 Max (xHigh reasoning)
0
c17f3b50-fc2c-4a9f-adbc-085a20c39ec0
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
50.6
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
45.6
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
33.9
%
GPT 6.1 Sol (Max)
OpenAI
32.2
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
31.1
%
GPT 6 Astra (Max)
OpenAI
16.7
%
Kimi K3.1 Pro (Max)
Moonshot AI
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
42.5
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
40.5
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
38.8
%
Muse Wonder 1.0 (Max)
Meta
37.3
%
GPT 6 Astra (Max)
OpenAI
36.5
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
33.5
%
GPT 6.1 Sol (Max)
OpenAI