Benchmarks
Agent37 on open sandbox benchmarks
We ran two public sandbox benchmarks with their own test harnesses. Here is how we compare on price and build speed, and all the raw data is public.
Cost of 1,000 builds
Clone, install and typecheck a real repo. Lower is better.
- Agent37$0.99
- Blaxel$9.81
- E2B$13.96
- Vercel Sandbox$27.64
- Modal$27.99
- Cloudflare Sandbox$37.98
Source: ComputeSDK benchmarks
Where we rank
Among every provider tested. #1 is best.
- Disk random reads · VM#1 of 17
- Disk sequential writes · default sandbox#1 of 17
- Hourly price · default sandbox#1 of 9
- Disk random writes · VM#3 of 17
- Installing dependencies · VM#4 of 17
- HTTPS to GitHub · default sandbox#4 of 16
- Building a real app · VM#5 of 17
- Node.js CPU · VM#7 of 17
Source: HPC Sandbox Benchmarks
HPC Sandbox Benchmarks
CPU per dollar and real-world builds
Median of fresh sandboxes at 4 vCPU and 8 GB RAM, ranked against every provider in the benchmark's latest published run. Close results share a rank.
CPU per dollar
Node.js runs per second divided by the hourly price, for every provider the benchmark prices.
| Provider | Runs/s | $ per hour | Runs/s per $1/hour |
|---|---|---|---|
| Agent37 | 17.53 | $0.0121 | 1,454 |
| boat | 28.06 | $0.0360 | 779.4 |
| Agent37 VM | 18.94 | $0.0301 | 628.5 |
| tama | 18.69 | $0.0740 | 252.6 |
| Novita | 20.16 | $0.233 | 86.42 |
| Daytona | 21.52 | $0.331 | 64.99 |
| Freestyle | 10.22 | $0.264 | 38.64 |
| E2B | 11.08 | $0.331 | 33.45 |
| Runloop | 8.93 | $0.634 | 14.09 |
Real-world build (Better-Auth)
| Metric | Default sandbox | VM | Best other provider |
|---|---|---|---|
| BuildSeconds, lower is better | 99.8#12 of 17 | 65.91#5 of 17 | Brezel42.23 |
| Cold installSeconds, lower is better | 24.12#13 of 17 | 14.42#4 of 17 | Brezel8.094 |
| Git cloneSeconds, lower is better | 8.305#16 of 17 | 1.768#9 of 17 | Namespace0.573 |
| Lint (Biome)Seconds, lower is better | 8.281#16 of 17 | 3.152#5 of 17 | Brezel2.2 |
| Lint deps (Knip)Seconds, lower is better | 15.58#12 of 17 | 10.49#3 of 17 | Brezel6.851 |
| Lint formatSeconds, lower is better | 3.741#5 of 17 | 3.109#5 of 17 | Brezel2.023 |
| Lint packagesSeconds, lower is better | 5.463#14 of 17 | 2.737#6 of 17 | Brezel1.774 |
| Lint spellSeconds, lower is better | 8.988#10 of 17 | 7.672#6 of 17 | Brezel4.865 |
| Lint typesSeconds, lower is better | 74.22#14 of 17 | 38.01#10 of 17 | Brezel20.62 |
| TypecheckSeconds, lower is better | 48.99#8 of 17 | 46.37#8 of 17 | Brezel27.57 |
Raw results and steps to reproduce: agent37-platform/hpc-sandbox-benchmarks, a copy of starslingdev/hpc-sandbox-benchmarks.
ComputeSDK benchmarks
Price and build time
ComputeSDK's pricing board and build benchmark, plus the cost of one build at each provider's own rate.
Price per hour
1 vCPU and 2 GB RAM, or a provider's smallest size when that is larger (2 vCPU and 4 GB for Agent37). ~ marks providers that bill for active CPU, which ComputeSDK prices at 10% use.
| Rank | Provider | $ per hour |
|---|---|---|
| 1 | Agent37 | $0.0060 |
| 2 | Blaxel | ~$0.0088 |
| 3 | Agent37 VM | $0.0151 |
| 4 | Agent37 Performance | $0.0241 |
| 5 | Cloudflare | ~$0.0252 |
| 6 | Vercel | ~$0.0340 |
| 7 | Modal | ~$0.0661 |
| 8 | CodeSandbox | $0.0743 |
| 9 | E2B | $0.0828 |
| 9 | Hopx | $0.0828 |
| 9 | Daytona | $0.0828 |
| 12 | Runloop | $0.1584 |
Build time
Clone, install and typecheck the OpenCode repo in a fresh sandbox. Median of successful builds.
| Rank | Provider | Median build | Builds | Size |
|---|---|---|---|---|
| 1 | isorun | 36.1 s | 1/1 | 8 vCPU / 16 GB |
| 2 | Promigence | 36.2 s | 1/1 | 8 vCPU / 32 GB |
| 3 | Namespace | 46.2 s | 1/1 | 8 vCPU / 16 GB |
| 4 | Blaxel | 53.0 s | 1/1 | 8 vCPU / 16 GB |
| 5 | CreateOS | 53.2 s | 1/1 | 8 vCPU / 16 GB |
| 6 | Upstash | 57.3 s | 1/1 | 48 vCPU / 185 GB |
| 7 | Mosaic | 58.6 s | 1/1 | 8 vCPU / 16 GB |
| 8 | GravixLayer | 58.7 s | 1/1 | 8 vCPU / 16 GB |
| 9 | Buddy | 59.1 s | 1/1 | 8 vCPU / 16 GB |
| 10 | Tenki | 60.8 s | 1/1 | 8 vCPU / 16 GB |
| 11 | Superserve | 64.3 s | 1/1 | 8 vCPU / 16 GB |
| 12 | Arker | 64.7 s | 1/1 | 8 vCPU / 16 GB |
| 13 | Agent37 Performance | 67.1 s | 3/5 | 4 vCPU / 16 GB |
| 14 | Sail | 67.3 s | 1/1 | 8 vCPU / 16 GB |
| 15 | Agent37 VM | 67.5 s | 5/5 | 4 vCPU / 16 GB |
| 16 | E2B | 75.9 s | 1/1 | 8 vCPU / 16 GB |
| 17 | Novita | 81.1 s | 1/1 | 8 vCPU / 16 GB |
| 18 | Vercel | 83.4 s | 1/1 | 8 vCPU / 16 GB |
| 19 | Beam | 96.1 s | 1/1 | 64 vCPU / 16 GB |
| 20 | Modal | 105.5 s | 1/1 | 4 vCPU / 329 GB |
| 21 | sandbox0 | 125.0 s | 1/1 | 8 vCPU / 16 GB |
| 22 | Agent37 | 133.5 s | 4/5 | 4 vCPU / 16 GB |
| 23 | Cloudflare | 189.9 s | 1/1 | 8 vCPU / 16 GB |
Cost per build
Median build time times each provider's hourly rate at the size it ran. Agent37 bills by the minute, so its builds are rounded up to whole minutes.
| Rank | Provider | Build | $ per hour | $ per build |
|---|---|---|---|---|
| 1 | Agent37 | 133 s | $0.0197 | $0.00099 |
| 2 | Agent37 VM | 68 s | $0.0493 | $0.00164 |
| 3 | Agent37 Performance | 67 s | $0.0789 | $0.00263 |
| 4 | Blaxel | 53 s | $0.6669 | $0.00981 |
| 5 | E2B | 76 s | $0.6624 | $0.01396 |
| 6 | Vercel | 83 s | $1.1936 | $0.02764 |
| 7 | Modal | 106 s | $0.9547 | $0.02799 |
| 8 | Cloudflare | 190 s | $0.7200 | $0.03798 |
Raw results and steps to reproduce: agent37-platform/computesdk-benchmarks, a copy of computesdk/benchmarks.
How to read these
- Other providers' results come from each benchmark's latest published run (October 3 and October 30, 2026). Ours ran on November 3, 2026, with the same harness and settings.
- Agent37 sandboxes go up to 4 vCPU. The build benchmark asks for 8, so our builds ran on half the cores most providers had.
- Two of five Performance builds ran out of memory and one standard build timed out. Medians count successful builds, and the Builds column shows every attempt.
- Our best disk results may include host caching. Read them as an upper bound.
- Prices are list rates with disk excluded. Agent37 is priced for the full time a sandbox runs.