After Being Called Out by Vercel's CEO, Agent37 Recalculated the Bill: 25x Becomes 14x
Vercel's CEO publicly said they got active CPU wrong. Agent37 reran a test with 100 agents, and the cheaper multiple shrank from 25x to 14x, also putting the cost of shared CPU mode on the table.
Agent37 previously said it was 25x cheaper than Vercel. Vercel CEO Guillermo Rauch replied on X: the calculation of active CPU time is completely wrong, creating a false cost comparison and a false impression.
Then Agent37 CEO Vishnu Krishnaprasad redid the math.
The method is not complicated: give 100 coding agents each a sandbox, have them vibe-code an app, and calculate the total hourly cost based on 5% active CPU. The results are as follows:
| Platform | Hourly cost |
|---|---|
| Vercel | $9.75 |
| E2B | $16.56 |
| Daytona | $16.56 |
| Agent37 | $0.68 |
14x cheaper, not 25x. The change in the number itself shows that the earlier claim was not rigorous enough, but the corrected result is still a huge gap.
Why is the difference so large? The operating rhythm of a coding agent is: call the model, wait for the response, read files, edit code, then call the model again. The CPU is actually busy only about 5% of the time; the rest is spent waiting on I/O and the model. Agent37 lets multiple agents share the same set of physical cores, replacing dedicated CPU for each agent with everyone taking turns using the CPU.
There is a detail here that is easy to skip over. In Vercel sandbox billing, active CPU is $0.128/vCPU-hour, and provisioned memory is $0.0212/GB-hour. For a 2 vCPU 4GB sandbox running for one hour at 5% active CPU, CPU cost is about $0.0128, and memory cost is about $0.0848. What drives the bill is mainly memory, not CPU. The Vercel CEO pointed out that you did not account for active CPU, but the cost gap is not entirely in that line item.
Someone in the comments pointed out the weakest part of the shared model: the average is 5%, until the moment 100 agents run test suites at the same time.
5% is a statistical average, not a real-time curve. When everyone is calling models, sharing is very cost-effective. When 100 agents simultaneously run builds, run tests, and install dependencies, CPU demand instantly stacks up. Agent37 dares to say it will refund the difference if you find a cheaper sandbox; it is betting that most users' workloads will not hit peak at the same time.
Agent37's pricing page also features a benchmark: running 1,000 builds costs $0.99, compared with $27.64 for Vercel Sandbox. The test code is open on GitHub and runs clone, install, and typecheck—I/O-heavy work, exactly the scenario where their model has the biggest advantage.
The low price enabled by shared CPU comes with preconditions. Now that the bill is clear, you also need to think through peak load.
发布时间: 2026-11-17 09:29