Replicate throttles harder when credit drops below $20: Sume 402
Replicate tightens limits as credit falls and advises keeping $20. Sume refuses a job with 402 insufficient_credits before provider work starts.

Replicate lowers your API rate limits as your account credit shrinks and says a balance above $20 with auto-reload avoids that; Sume instead refuses a job outright with 402 insufficient_credits before provider work starts. One slows you down near zero, the other stops you at the submit call.
Replicate's rules are on its rate limits page; Sume's are in Errors and rate limits and Generation admission, read 2026-10-02.
What are Replicate's numbers?
Replicate lists 600 requests per minute for creating predictions and 3000 per minute for all other endpoints, with short bursts allowed before throttling. A throttled call gets a 429 with a message that the rate limit resets in about 30 seconds.
The credit rule is the unusual part: as your credit decreases, stricter limits apply automatically, to prevent accidental overspending. Accounts with granted credits but no payment method get 1 request per second, capped at 6 per minute. Higher limits need a support request.
What does Sume do near zero balance?
Sume checks spendable USD balance at admission. If it cannot reserve the estimated cost, the submit fails with 402 insufficient_credits before provider work starts. Its docs tell you to upgrade the plan, wait for included Gen$, or submit a cheaper request, and not to invent prepaid top-ups.
Nothing on these pages says Sume lowers a request rate as balance falls. A low balance does not slow your calls; it only rejects the one that cannot be reserved.
| Question | Replicate | Sume |
|---|---|---|
| Balance falls | Limits tighten automatically | Same limits; submit refused if estimate cannot be reserved |
| Error | 429 throttled | 402 insufficient_credits |
| Retry advice | Wait about 30 s; keep over $20 | Cheaper request or more balance, not a blind retry |
| Cost known before submit | Per-model pricing | USD estimate reserved at accept |
How should a batch job handle each?
On Replicate, a batch near a low balance can look like a rate-limit bug. Check the balance before blaming concurrency. On Sume, a 402 is not retryable by waiting a few seconds, so route it to a human or a smaller request rather than the 429 backoff loop.
Sume exposes /v1/balance and /v1/usage so you can read the balance and ledger before a large run. Failed jobs and failed queue admission release or refund the reservation where applicable, so a 402 on one job does not leave stray holds from the others.
Does Sume have a free tier like Replicate's granted credits?
These pages do not describe a granted-credit rate cap for Sume. Sume's plans and included Gen$ are covered on its pricing pages; for how a first run is billed, read the Core workflow usage section before you size a batch.
Sources
Related posts
More in Pricing
- Runway Gen-4 Turbo and Wan3 Prime cost per second in dollars
At $0.01 per credit, Runway Gen-4 Turbo is $0.05 per second and Wan3 Prime is $0.068, $0.14 and $0.28 at 480p, 720p and 1080p. How to compare on Sume.
- Grok Imagine Video 1.5 on Runway: 10, 16, 29 credits a second
Runway bills grok_imagine_1_5 at 10, 16 or 29 credits per second by resolution, plus 1 per reference. Sume takes grok-imagine-video-1.5 as image-to-video.
- Runway real-time avatar cost per minute vs a rendered avatar video
Runway bills real-time avatars at 2 credits upfront plus 2 per 6 seconds, about $0.22 a minute. Sume's rendered avatar video is $0.184 to $0.55 per second.
- Seedance 2.5 token pricing on fal: formula and a 30-second 720p clip
fal bills Seedance 2.5 as height x width x seconds x 24 / 1024 tokens at $0.0214 per 1,000. Worked examples, a rounding gap on the page, and what Sume lists.
Written by Sume