DAYJOB: Law
Professional Work
Surge AI's DAYJOB: Law extends the DAYJOB suite of long-horizon professional assignments to legal work: each task is a request as brief as a colleague's message plus a workspace of the files it refers to, and an agent judge applies an expert-written rubric to the delivered work. An attempt passes only when it meets at least 95% of the rubric criteria, and the headline is pass@1 at that threshold, as on Surge's DAYJOB: Tax board. Surge runs the benchmark and supplied its first results to Fireworks' Specialized Intelligence Index before publishing the tasks, their number, the harness or the attempt count; the DAYJOB family's published harness is the OpenHands SDK in Harbor with no browsing or package downloads.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 36.9 |
| Muse Wonder | 36.6 |
| GPT-6.1 Sol | 31.8 |
| Grok 4.7 | 27.6 |
| Claude Sonnet 5.5 | 27.0 |
| Claude Fable 5.1 | 26.5 |
| Claude Opus 5.5 | 25.1 |
| Muse Spark 1.3 | 22.8 |
| Gemini 4 Argon | 9.9 |
| DeepSeek V4.1 Flash 1017 | 6.2 |
| GLM-5.3 | 6.2 |
| Mistral Large 4 Preview | 4.5 |
| Kimi K3 | 4.2 |