claude-opus-5 leads the current leaderboard.
Its Overall score is 95.6 points. The gap to gpt-5.6-sol is 6.0 points.
ReconBench Model Leaderboard
Overall score
| Rank | Model | Agent | Overall | Visual | Scored runs | Case coverage | Attempts | |
|---|---|---|---|---|---|---|---|---|
| 1 | claude-opus-5 | Claude Codehigh effort · baseline | 95.6 | 96.0 | 5 / 5 | 3 / 3 | 5 | |
| 2 | gpt-5.6-sol | Codexhigh effort · baseline | 89.7 | 90.3 | 5 / 6 | 3 / 3 | 6 | |
| 3 | gpt-5.6-terra | Codexhigh effort · baseline | 70.6 | 81.1 | 5 / 5 | 3 / 3 | 5 | |
| 4 | claude-sonnet-5 | Claude Codehigh effort · baseline | 67.9 | 70.9 | 5 / 5 | 3 / 3 | 5 | |
| 5 | gpt-5.6-luna | Codexhigh effort · baseline | 54.1 | 67.0 | 5 / 5 | 3 / 3 | 5 | |
Overall is the equal-weight macro-average across all 3 active cases. Only completed, scored attempts enter case means. Candidate failures, timeouts, and infrastructure failures remain in the audit record but are excluded from scores.
How to read the leaderboard
New cases update scores automatically.
Each case contributes equal weight, regardless of how many checkpoints or repeated attempts it contains.
Coverage gates the ranking.
5 of 5 model entries currently cover all 3 active cases. Partial entries remain visible but unranked.
Benchmark coverage
3active cases
5model entries
5fully ranked
26trusted attempts
| Active case | Track | Targets | States | Runs |
|---|---|---|---|---|
| Reconstruct a browser-side Base64 workbenchcyberchef-base64-workbench | reconstruct | 2 | 0 | 10 |
| Reconstruct a local image-compression comparison toolsquoosh-image-compression | reconstruct | 2 | 0 | 6 |
| Reconstruct a local-first Markdown workspacestackedit-field-notes | reconstruct | 2 | 1 | 10 |
Inspect all casesOpen run recordsRead the aggregation policy