Benchmarks
MIOSA in the independent ComputeSDK sandbox benchmarks: burst startup and DAX builds. Every result is a committed JSON file, so each number can be checked against the repository.
Where MIOSA stands
Burst startup
2026-10-23#1of 32
57.1 ms median
p95 68.81 ms · p99 72.20 ms · 100/100 succeeded
DAX build
best run#1Aug 21 & Aug 28
38.4 s total build in its #1 run
Latest run (2026-10-23): #3 of 22 completed at 38.8 s. Fastest MIOSA build so far: 36.5 s (2026-10-16, #2).
MIOSA does not appear in the other ComputeSDK suites (browser, storage, AI gateway, and the retired sequential and staggered TTI tests), because it does not run those workloads.
Rank history
Burst startup
DAX build
Ranking is by the metric the suite uses at the top of its leaderboard: composite score for burst TTI (rewards speed and reliability together), and total build time for DAX. A provider with 0% success has no result to rank. On the other DAX runs MIOSA was #2, #3, #4 or #18, and on three dates the run failed.
Leaderboard
Burst TTI, 2026-10-23, all 32 providers
The latest run. Ranked by composite score. Success is out of 100 iterations. MIOSA's row is highlighted. 100 concurrent sandbox create requests, each timed from the API request to the first successful command inside the fresh sandbox. Lower is better.
| Rank | Provider | Median (ms) | Relative bar | p95 (ms) | p99 (ms) | Success |
|---|---|---|---|---|---|---|
| 1 | miosa | 57.09ms | 68.81ms | 72.2ms | 100% | |
| 2 | isorun | 74.93ms | 99.3ms | 103.47ms | 100% | |
| 3 | promigence | 91.75ms | 106.63ms | 108.06ms | 100% | |
| 4 | gravixlayer | 108.71ms | 116.39ms | 116.99ms | 100% | |
| 5 | archil | 189.87ms | 251.19ms | 269.15ms | 100% | |
| 6 | arker | 316.5ms | 335.32ms | 343.75ms | 100% | |
| 7 | blaxel | 441.82ms | 476.2ms | 484.62ms | 100% | |
| 8 | superserve | 410.28ms | 559.81ms | 638.99ms | 100% | |
| 9 | createos | 466.18ms | 576.2ms | 578.24ms | 100% | |
| 10 | givemeanode | 501.73ms | 694.41ms | 696.06ms | 100% |
Two providers (hopx, northflank) recorded no successful iteration, so they have no median to rank. Bars show median on a log scale.
Burst startup over time
703.63 ms2026-09-04
57.09 ms2026-10-23
7 weeks| Date | Median (ms) | Rank |
|---|---|---|
| 2026-08-28 | 698.6 | #9 of 29 |
| 2026-09-04 | 703.6 | #7 of 31 |
| 2026-09-11 | 222 | #3 of 29 |
| 2026-09-16 | 500.2 | #8 of 30 |
| 2026-09-18 | 399.1 | #18 of 29 |
| 2026-09-25 | 361.2 | #4 of 30 |
| 2026-10-02 | 163.4 | #2 of 30 |
| 2026-10-09 | 117.9 | #1 of 30 |
| 2026-10-16 | 94.8 | #2 of 31 |
| 2026-10-20 | 73.1 | #2 of 32 |
| 2026-10-23 | 57.1 | #1 of 32 |
Method and reproduction
How it's run
Independent and open source
The benchmark code, runner and every result file are public. ComputeSDK states that sponsors cannot influence results. github.com/computesdk/benchmarks
Same code, every provider
The same open code against every provider, with fixed workloads and a fixed scoring ceiling. Each metric reports min, max, median, p95 and p99, and burst TTI also reports a composite score that rewards speed and reliability together.
Automated
Runs execute on GitHub Actions on Namespace runners, and every result is committed to the repository as JSON. The repository lists the schedule as daily at 5pm Pacific, 00:00 UTC. In practice the burst TTI history is daily through 2026-07-13 and then weekly on Fridays, with extra runs on 2026-09-10 and 2026-09-16.
Live leaderboards and methodology
Live leaderboards and the full methodology are published by ComputeSDK. computesdk.com/benchmarks
Reproduce it
git clone https://github.com/computesdk/benchmarks
cd benchmarks
# MIOSA's burst TTI row, latest run
jq '.results[] | select(.provider=="miosa")' results/burst_tti/2026-10-23.json
# Every provider, sorted by median
jq -r '.results[] | "\(.summary.ttiMs.median)\t\(.provider)"' results/burst_tti/2026-10-23.json | sort -nThe DAX overwrite is visible in history:
git log --oneline -- results/sandbox-dax/2026-10-09.json
git show 8120d15b:results/sandbox-dax/2026-10-09.json | jq '.results[] | select(.provider=="miosa")'