How the estimates work.
A million non-overlapping calls, each followed by the selected idle gap before the next call. Agent37, Boat and Cloudflare rates and assumptions were checked November 5, 2026; other rates use the site’s October 12–19 snapshot. This uses unrounded list-rate arithmetic, not invoice billing units, and estimates resource charges rather than performance or a complete monthly bill.
Included and excluded.
Included: execution, billed idle time between requests, database operations and outbound bytes at list rates. Excluded: subscriptions, plan allowances, free tiers, storage at rest, startup and wake overhead, upstream API charges and model tokens.
Calls run one at a time with equally spaced idle gaps. The estimate covers the first call’s start through the last call’s return, with 999,999 gaps. Zero idle means back-to-back calls. Agent37 and Boat stay awake across gaps; Deno bills memory until its modeled shutdown. Other rows assume no billable work between requests.
Concurrent calls can share allocations and reduce costs on some platforms; memory bars show the modeled allocation, not measured usage.
The worked SQLite index example separately applies CrabTalk Go and Workers Paid row allowances to a shared research library. Its assumed counters are not measured performance results.
Database rows, KV units and Redis commands are modeled equivalents with different capabilities and guarantees. Apify, Railway, Rivet and Google Cloud Run are outside this calculator; their pricing models are covered on their comparison pages.
Platform assumptions.
- CrabTalk
- The memory input is the tool’s budget, held until return. CPU beyond the run’s initial 0.1 ms is billed.
- Agent37 Standard
- One shared 2 vCPU / 4 GB sandbox, kept awake across calls and idle gaps at $0.80/vCPU-month and $0.70/GB-month (730 hours), with total uptime rounded to minutes. Database operations use local SQLite; disk and unquoted transfer costs are excluded. Optional auto-sleep can reduce awake time; this estimate uses the default with auto-sleep off.
- Boat Small
- One reused Small VM: 2 vCPU, 4 GB RAM and 51 GB disk at $0.018/hour, billed by the second through calls and idle gaps. Auto-stop is disabled for a continuous service. SQLite runs locally and egress is included. The $20 monthly plan minimum, credited toward VM time, is excluded; optional snapshots and excess starts are excluded too.
- Durable Objects + Worker
- One Worker request and one object request per call, with 128 MiB allocated. The object is eligible for hibernation after returning, with no timers or other work keeping it billable between requests. Worker CPU is excluded.
- Convex
- Professional action rates in a US region. The model selects the 64 MiB or 512 MiB runtime and transfers 1 KiB per row; index I/O is excluded.
- Deno Deploy
- Pro rates at the default 768 MiB allocation. One KV unit per row. Memory bills during calls and up to 20 seconds of each idle gap: the lower end of the documented roughly 20–30 seconds before idle apps shut down. Idle CPU is not charged.
- AWS Lambda + DynamoDB
- 128 MiB x86 in US East, with CPU duration modeled at roughly 1/14 of a vCPU. One strongly consistent read or write per row up to 1 KiB.
- Vercel + Upstash
- US Pro rates at 2 GiB provisioned memory. One Upstash command per row. CPU and memory charges pause after the last in-flight request completes. Concurrent calls can share the instance’s memory allocation.