Ox Alpha resolves 48 of 50 SWE-bench Verified Mini tasks

A user benchmarked the Ox Alpha model on the SWE-bench Verified-Mini dataset, achieving a 96% resolution rate (48 out of 50 tasks) using the official mini-swe-agent scaffold. The evaluation was conducted locally with the official SWE-bench Docker harness, where a task is considered resolved only if all FAIL_TO_PASS and PASS_TO_PASS tests pass.

• Model: Ox Alpha served through opencode's Go gateway.
• Scaffold: mini-swe-agent v2.4.6 (Bash-Only official leaderboard scaffold).
• Dataset: swe-bench-verified-mini, a 50-task subset containing only django and sphinx repositories.
• Results: 48/50 resolved, with failures in django__django-11790 and django__django-11815; average of 40 steps per task.
• Comparison: The score exceeds Claude Fable 5's 95% on the same subset, though the author notes this may be inflated due to memorization or a smaller, easier dataset.

The author considers the result suspiciously high for a free-tier model and advises treating the 96% figure as an upper bound rather than a comparable benchmark against models tested on the full 500-task set.