Methodology

Where the numbers come from

Each figure comes from the provider’s own pricing page, documentation or GitHub repository. Every record lists its sources with the date we read them, and every value links straight to the page it came from. Every number is the provider’s own published figure or plain arithmetic on two of them. Where a provider keeps a figure unpublished, the cell shows a dash and the provider page says “not published”.

How prices become $/vCPU-hour and $/GiB-hour

Providers publish rates per second, per millisecond, per hour, per core or per bundled unit. The table converts every rate it can to two units, so a row can be compared with any other row:

• Per-second rates × 3,600. E2B publishes $0.000014 per vCPU-second and $0.0000045 per GiB-second, which become $0.0504 per vCPU-hour and $0.0162 per GiB-hour (source: E2B Pricing). Cloudflare, Hopx and Blaxel are converted the same way.
• Per-core rates ÷ vCPUs per core, where the provider defines a core. Modal bills per physical core and defines one core as two vCPUs, so its $0.00003942 per core-second (its non-preemptible rate, which sandboxes pay) becomes $0.142 per core-hour and $0.071 per vCPU-hour (source: Modal Pricing). Where a provider prices per “CPU” or “core” without defining it as more than one vCPU (Runloop, Tensorlake, Fly.io Sprites), one core counts as one vCPU.
• GB is read as GiB where the provider does not say which it means. The two differ by 7%.
• Memory free means $0. Upstash Box states memory is not billed, so its memory cell is $0, not a dash.
• Bundled units stay out of the per-vCPU column. Morph sells Morph Compute Units: one MCU ($0.05) buys 1 vCPU-hour, 4 GB RAM-hours and 16 GB disk-hours. That is listed as a sandbox-hour price, not split into rates Morph does not publish.
• Lowest public tier, default region, on-demand. Committed-use discounts, enterprise quotes and plan credits are left out. Vercel’s rates are its iad1 rates.

Wall-clock vs active-CPU billing

Most providers bill every second a sandbox is allocated (wall clock). Vercel, Tensorlake and Upstash Box bill CPU only while it is busy; Blaxel bills memory while a sandbox is active and nothing in standby. The Billing basis column says which applies. For an agent that waits on a model between tool calls, an active-CPU rate that looks higher can cost less per task.

The 2 vCPU / 4 GiB box

The E2B alternatives page and each provider page price one sandbox with 2 vCPUs and 4 GiB of memory for one hour: 2 × the vCPU rate + 4 × the memory rate. It is computed only for managed providers that publish both rates, and disk is left out because most include a free allowance. For active-CPU providers it is the price at full load.

Start times and limits

Cold-start and resume times are the provider’s own published numbers, with their wording and percentile in the Startup claim field. 5 providers publish a cold-start figure, and the column lists exactly those. Session length, concurrency and machine-size limits are taken per plan as published.

Checks and corrections

Every row carries the date we last checked it; the newest is 9 Nov 2026. Providers and readers can send a correction with a public source to hello@sandboxproviders.com. We update the row and its date. Sort order comes from published prices alone; the paid featured slot is labelled and sits under the table.