DAYJOB: Tax A benchmark for AI tax agents

Blog
benchmarks
Off-the-Shelf
Training
experts
Products
Enterprise
Research
Careers
Contact
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.

Blog
Research
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.
Riemann-bench Leaderboard
1
GPT
5.6 Sol (Max reasoning)
83.2
%
1
GPT
6 Astra (Max reasoning)
81.6
%
1
Gemini
4 Carbon (High reasoning)
80
%
1
Claude
Opus 5.5 (Adaptive/Max)
80
%
1
GPT
6.1 Sol (Max reasoning)
79.2
%
1
Claude
Sonnet 5.5 (Adaptive/Max)
76.8
%
1
Gemini
4 Argon (High reasoning)
74.4
%
1
Claude
Fable 5.1 (Adaptive/Max)
71.2
%
1
Muse
Wonder Full (Max reasoning)
63.2
%
1
Kimi
K3.1 Pro (Max reasoning)
60.8
%
1
Muse
Wonder 1.0 (Max reasoning)
59.2
%
1
Grok
4.7 (xHigh reasoning)
57.6
%
1
Gemini
3.8 Flash (High reasoning)
56.8
%
1
Muse
Spark 1.3 (xHigh reasoning)
53.6
%
1
Qwen
3.8 Max (xHigh reasoning)
48
%
1
DeepSeek
V4.1 Flash 1017 (Max reasoning)
38.4
%
1
Beam
1.0 (Max reasoning)
37.6
%
1
Kimi
K3 (Max reasoning)
36.8
%
1
Hy
Hy4 Preview (High reasoning)
24
%
1
GLM
5.3 (Max reasoning)
20.8
%
1
Mistral
Large 4 (High reasoning)
20
%

Measuring progress along the mathematical frontier.

Read the blog post
*
*      *

More benchmarks

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.

Leaderboard
Blog
Hugging Face
Rank
Model
Score
Claude Sonnet 5.5 (Adaptive/Max)
10.8
%
Claude Sonnet 5.5 (Adaptive/Max)
Claude Fable 5.1 (Adaptive/Max)
9.6
%
Claude Fable 5.1 (Adaptive/Max)
Claude Opus 5.5 (Adaptive/Max)
9.2
%
Claude Opus 5.5 (Adaptive/Max)
Muse Wonder 1.0 (Max)
6
%
Muse Wonder 1.0 (Max)
Grok 4.7 (xHigh)
5
%
Grok 4.7 (xHigh)
GPT 6 Astra (Max)
4.8
%
GPT 6 Astra (Max)
GPT 6.1 Sol (Max)
4
%
GPT 6.1 Sol (Max)
Muse Spark 1.3 (Max)
3.6
%
Muse Spark 1.3 (Max)
Gemini 4 Argon (High)
3.2
%
Gemini 4 Argon (High)
Hy4 Preview
1.6
%
Hy4 Preview
DeepSeek V4.1 Flash 1017 (Max)
1.2
%
DeepSeek V4.1 Flash 1017 (Max)
Gemini 3.8 Flash (High)
0.8
%
Gemini 3.8 Flash (High)
GLM 5.3 (Max)
0.4
%
GLM 5.3 (Max)
Kimi K3 (Max)
0
%
Kimi K3 (Max)
Mistral Large 4 (High)
0
%
Mistral Large 4 (High)
Qwen 3.8 Max
0
%
Qwen 3.8 Max
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)
50.6
%
Claude Sonnet 5.5 (Adaptive/Max)
Claude Opus 5.5 (Adaptive/Max)
45.6
%
Claude Opus 5.5 (Adaptive/Max)
GPT 6.1 Sol (Max)
33.9
%
GPT 6.1 Sol (Max)
Claude Fable 5.1 (Adaptive/Max)
32.2
%
Claude Fable 5.1 (Adaptive/Max)
GPT 6 Astra (Max)
31.1
%
GPT 6 Astra (Max)
Kimi K3.1 Pro (Max)
16.7
%
Kimi K3.1 Pro (Max)
Qwen 3.8 Max
16.7
%
Qwen 3.8 Max
Muse Wonder 1.0 (Max)
16.1
%
Muse Wonder 1.0 (Max)
Gemini 4 Argon (High)
16.1
%
Gemini 4 Argon (High)
Hy4 Preview
15.6
%
Hy4 Preview
Inkling 2 Small (xHigh)
14.1
%
Inkling 2 Small (xHigh)
Mistral Large 4 (High)
13.9
%
Mistral Large 4 (High)
Muse Spark 1.3 (Max)
12.8
%
Muse Spark 1.3 (Max)
Grok 4.7 (xHigh)
12.2
%
Grok 4.7 (xHigh)
Kimi K3 (Max)
12.2
%
Kimi K3 (Max)
Gemini 3.8 Flash (High)
10.6
%
Gemini 3.8 Flash (High)
GLM 5.3 (Max)
9.4
%
GLM 5.3 (Max)
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)
Grok 4.7 (xHigh)
22.3
%
Grok 4.7 (xHigh)
GPT 6.1 Sol (Max)
22
%
GPT 6.1 Sol (Max)
GLM 5.3 (Max)
21.4
%
GLM 5.3 (Max)
Qwen 3.8 Max
20.3
%
Qwen 3.8 Max
Hy4 Preview
18.3
%
Hy4 Preview
Gemini 3.8 Flash (High)
18.3
%
Gemini 3.8 Flash (High)
GPT 6 Sol (Max)
17.1
%
GPT 6 Sol (Max)
Claude Sonnet 5 (Adaptive/Max)
16.9
%
Claude Sonnet 5 (Adaptive/Max)
GPT 6 Luna (Max)
14.9
%
GPT 6 Luna (Max)
Kimi K3 (Max)
13.4
%
Kimi K3 (Max)
Gemini 3.1 Pro
10.3
%
Gemini 3.1 Pro
DeepSeek V4 Pro (Max)
6.9
%
DeepSeek V4 Pro (Max)
Nemotron 3 Ultra
4.9
%
Nemotron 3 Ultra
Mistral Large 3
0
%
Mistral Large 3
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)
48.3
%
Claude Opus 5.5 (Adaptive/Max)
Claude Sonnet 5.5 (Adaptive/Max)
34
%
Claude Sonnet 5.5 (Adaptive/Max)
Claude Fable 5.1 (Adaptive/Max)
31.2
%
Claude Fable 5.1 (Adaptive/Max)
Muse Wonder 1.0 (Max)
31.2
%
Muse Wonder 1.0 (Max)
GPT 6 Astra (Max)
28.4
%
GPT 6 Astra (Max)
Grok 4.7 (xHigh)
27.6
%
Grok 4.7 (xHigh)
GPT 6.1 Sol (Max)
25.2
%
GPT 6.1 Sol (Max)
Gemini 4 Argon (High)
23.6
%
Gemini 4 Argon (High)
Muse Spark 1.3 (Max)
22
%
Muse Spark 1.3 (Max)
Claude Opus 5 (Adaptive/Max)
22
%
Claude Opus 5 (Adaptive/Max)
Grok 4.6 (xHigh)
14.8
%
Grok 4.6 (xHigh)
Claude Fable 5 (Adaptive/Max)
11.2
%
Claude Fable 5 (Adaptive/Max)
Muse Spark 1.2 (xHigh)
8.8
%
Muse Spark 1.2 (xHigh)
GPT 6 Sol (Max)
7.2
%
GPT 6 Sol (Max)
Hy4 Preview
6.4
%
Hy4 Preview
DeepSeek V4.1 Flash 1017 (Max)
6
%
DeepSeek V4.1 Flash 1017 (Max)
Qwen 3.8 Max
4.8
%
Qwen 3.8 Max
GLM 5.3 (Max)
4.4
%
GLM 5.3 (Max)
GPT 5.6 Sol (Max)
4.4
%
GPT 5.6 Sol (Max)
Kimi K3 (Max)
4
%
Kimi K3 (Max)
GPT 5.6 Terra (xHigh)
3.6
%
GPT 5.6 Terra (xHigh)
Mistral Large 4 (High)
3.6
%
Mistral Large 4 (High)
Claude Sonnet 5 (Adaptive/Max)
2.8
%
Claude Sonnet 5 (Adaptive/Max)
Gemini 3.7 Flash (High)
2.4
%
Gemini 3.7 Flash (High)
Gemini 3.8 Flash (High)
2
%
Gemini 3.8 Flash (High)
GPT 5.6 Luna (Max)
1.6
%
GPT 5.6 Luna (Max)
GLM 5.3 Flash (Max)
1.2
%
GLM 5.3 Flash (Max)
DeepSeek V4 Pro (Max)
1.2
%
DeepSeek V4 Pro (Max)
Hy3 (High)
1.2
%
Hy3 (High)
GPT 6 Luna (Max)
1.2
%
GPT 6 Luna (Max)
Reflection Beam 1.0 (Max)
1.2
%
Reflection Beam 1.0 (Max)
Inkling (Max)
0.4
%
Inkling (Max)
Gemini 3.1 Pro
0.4
%
Gemini 3.1 Pro
Nemotron 3 Ultra
0.4
%
Nemotron 3 Ultra
DeepSeek V4 Flash (Max)
0
%
DeepSeek V4 Flash (Max)
Kimi K2.7 Code (Max)
0
%
Kimi K2.7 Code (Max)
Muse Glimmer 30B (xHigh)
0
%
Muse Glimmer 30B (xHigh)
Mistral Large 3
0
%
Mistral Large 3
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)
42.5
%
Claude Opus 5.5 (Adaptive/Max)
Claude Sonnet 5.5 (Adaptive/Max)
40.5
%
Claude Sonnet 5.5 (Adaptive/Max)
Muse Wonder 1.0 (Max)
38.8
%
Muse Wonder 1.0 (Max)
GPT 6 Astra (Max)
37.3
%
GPT 6 Astra (Max)
Claude Fable 5.1 (Adaptive/Max)
36.5
%
Claude Fable 5.1 (Adaptive/Max)
GPT 6.1 Sol (Max)
33.5
%
GPT 6.1 Sol (Max)
Gemini 4 Argon (High)
30.5
%
Gemini 4 Argon (High)
Muse Spark 1.3 (Max)
27
%
Muse Spark 1.3 (Max)
Grok 4.7 (xHigh)
26.5
%
Grok 4.7 (xHigh)
Claude Opus 5 (Adaptive/Max)
23.8
%
Claude Opus 5 (Adaptive/Max)
Grok 4.6 (xHigh)
21
%
Grok 4.6 (xHigh)
GPT 6 Sol (Max)
19.5
%
GPT 6 Sol (Max)
Claude Fable 5 (Adaptive/Max)
18.3
%
Claude Fable 5 (Adaptive/Max)
Kimi K3.1 Pro (Max)
17
%
Kimi K3.1 Pro (Max)
Inkling 2 Small (xHigh)
14.3
%
Inkling 2 Small (xHigh)
DeepSeek V4.1 Flash 1017 (Max)
12.5
%
DeepSeek V4.1 Flash 1017 (Max)
GPT 5.6 Sol (Max)
11.8
%
GPT 5.6 Sol (Max)
GLM 5.3 (Max)
11
%
GLM 5.3 (Max)
Hy4 Preview
9.3
%
Hy4 Preview
Qwen 3.8 Max
8.8
%
Qwen 3.8 Max
GLM 5.3 Flash (Max)
7.3
%
GLM 5.3 Flash (Max)
Muse Spark 1.2 (xHigh)
6.8
%
Muse Spark 1.2 (xHigh)
GPT 6 Luna (Max)
6.8
%
GPT 6 Luna (Max)
Kimi K3 (Max)
5.5
%
Kimi K3 (Max)
Gemini 3.7 Flash (High)
5.3
%
Gemini 3.7 Flash (High)
Mistral Large 4 (High)
4.8
%
Mistral Large 4 (High)
Claude Sonnet 5 (Adaptive/Max)
4.5
%
Claude Sonnet 5 (Adaptive/Max)
Gemini 3.8 Flash (High)
4.5
%
Gemini 3.8 Flash (High)
GPT 5.6 Luna (Max)
4.3
%
GPT 5.6 Luna (Max)
DeepSeek V4 Pro (Max)
3.5
%
DeepSeek V4 Pro (Max)
GPT 5.6 Terra (xHigh)
3.3
%
GPT 5.6 Terra (xHigh)
Reflection Beam 1.0 (Max)
1.5
%
Reflection Beam 1.0 (Max)
Hy3 (High)
0.5
%
Hy3 (High)
DeepSeek V4 Flash (Max)
0.3
%
DeepSeek V4 Flash (Max)
Gemini 3.1 Pro
0.3
%
Gemini 3.1 Pro
Inkling (Max)
0
%
Inkling (Max)
Kimi K2.7 Code (Max)
0
%
Kimi K2.7 Code (Max)
Mistral Large 3
0
%
Mistral Large 3
Muse Glimmer 30B (xHigh)
0
%
Muse Glimmer 30B (xHigh)
Nemotron 3 Ultra
0
%
Nemotron 3 Ultra

Stay Posted

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