Legacy-Bench (Fireworks, avg@3)
Code
Factory's Legacy-Bench tests software-engineering agents on 100 maintenance tasks in legacy stacks (COBOL, Java 7, BASIC, C89, Fortran, and Assembly): fixing bugs, implementing features, and migrating code, each in a containerized Harbor environment with hidden verification tests. Fireworks runs all 100 tasks three times per model with Harbor's terminus-2 agent in E2B sandboxes, at max reasoning effort with tripled task timeouts. The score is the mean over three runs of the share of tasks whose hidden tests all pass; it is not comparable with Factory's own Droid-harness runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 62.0 |
| Muse Spark 1.3 | 62.0 |
| DeepSeek V4.1 Flash | 57.7 |
| DeepSeek V4.1 Flash 1017 | 56.3 |
| GPT-6 Astra | 55.7 |
| GPT-6.1 Sol | 55.0 |
| Grok 4.7 | 54.0 |
| Kimi K3 | 52.7 |
| Gemini 3.8 Flash | 52.3 |
| GLM-5.3 | 50.7 |
| Muse Spark Open | 50.0 |
| Gemini 4 Argon | 44.3 |