Supabase Evals
2026-10-10 · Framework eval · 23 tasks × 3 runs
- What it measures
- Models do auth, database, RLS, storage, and debugging work on Supabase. 23 tasks, 3 runs each, 69 checks.
- How
- The official Supabase evals. The baseline is no-skills: the supabase and supabase-postgres-best-practices skills stay unloaded. The other arm loads them. The two columns stay separate and are not merged into one pass rate.
- What the score means
- The percent is passes out of 69 checks. Rank follows the no-skills baseline. The change is with-skills minus baseline, in percentage points, and it can be zero or negative.
- Caveats
- Supabase writes the tasks and grades them. This measures a model together with Supabase, not one framework against another. Six models and 23 tasks. The vendor authors the tasks and grades them.
- Updates
- Aggregated on 2026-10-10 from the supabase/evals results. The file has no separate commit version, so the version is recorded as 23 tasks × 3 runs.
Baseline board
| # | Model | Baseline |
|---|---|---|
| 1 | Claude Opus 5.5 high · Claude Code | 91.3 |
| 2 | Kimi K3 OpenCode | 84.1 |
| 3 | Grok 4.7 high · Grok | 78.3 |
| 4 | GPT 6 Sol medium · Codex | 78.3 |
| 5 | Claude Sonnet 5 high · Claude Code | 75.4 |
| 6 | GPT 6 Luna medium · Codex | 63.8 |