Benchmark evidence
A leaderboard that explains itself.
Compare center, tail, failure rate, cost and sample size. Rankings change with the workload; missing evidence stays visible instead of becoming a synthetic score.
Fastest median
895 ms
Tensorlake
Tightest tail
1.05×
Tensorlake · p99 / p50
Most reliable
100.0%
Blaxel
Lowest run cost
$0.00000
Tensorlake
Explore the evidence
Fastest median wall time first.
| # | Provider | Median | p95 | p99 | Success | Cost | n | Signal |
|---|---|---|---|---|---|---|---|---|
| 01 |  Tensorlake | 895 ms | — | 939 ms | 100.0% | $0.00000 | 3 | Early |
| 02 |  E2B | 967 ms | — | 11.13 s | 92.7% | $0.00009 | 151 | High |
| 03 |  blaxel | 1.19 s | — | 2.45 s | 100.0% | $0.00002 | 43 | Medium |
| 04 |  declaw | 1.83 s | — | 3.95 s | 100.0% | $0.00003 | 52 | Medium |
| 05 |  Daytona | 2.05 s | — | 25.40 s | 100.0% | $0.00007 | 93 | Medium |
| 06 |  Vercel | 3.05 s | — | 17.46 s | 85.4% | $0.00039 | 103 | High |
| 07 |  CodeSandbox | 5.09 s | — | 14.13 s | 97.7% | $0.00008 | 43 | Medium |
| 08 |  Dedalus | 6.29 s | — | 28.48 s | 92.3% | $0.00009 | 130 | High |
| 09 |  Sprites | 15.50 s | — | 33.89 s | 91.3% | $0.00033 | 126 | High |
| — |  CLOUDFLARE | Untested | — | — | — | — | — | None |
| — |  hopx | Untested | — | — | — | — | — | None |
| — |  Modal | Untested | — | — | — | — | — | None |
| — |  Northflank | Untested | — | — | — | — | — | None |
| — |  RUNLOOP | Untested | — | — | — | — | — | None |
Missing percentiles stay empty—no synthetic fallback744 measured runs in view
Read the middle
Median is the normal experience. Use it for interactive startup and repeatable tool calls.
Read the tail
p95 and p99 expose jitter, queueing and slow recovery paths that averages hide.
Read the sample
Under 30 runs is directional. At 100+ runs the percentile shape becomes substantially more useful.