elevenlabs.providers.sgit.ai / experiments / Concurrency probe — where is the wall?
Concurrency probe — where is the wall?
Concurrent request limits scale with the plan, and a render that fires every scene at once — 10 to 16 requests — meets that limit before it meets any other. We have never provoked one. This lab does, deliberately, for a few cents.
Your own full key, in your own browser, bounded only by your plan's monthly quota. That is pattern 0 with a ceiling: the narrow case where it is defensible — the key's owner, testing their own key, on their own device — and not a pattern to publish. No key ships in this page: the field below is empty until you fill it.
ElevenLabs offers no per-key spend limit for text to speech, so nothing here caps what a leaked key could cost you except the account's own quota. Use a key scoped in the dashboard to the endpoints this lab needs, and forget it when you are done. The version of this page with no key box at all is pattern three, and it does not exist yet.
This is the one lab here that can cost more than pennies. It generates real audio, in parallel, until something breaks. The estimate above the button is the worst case for the numbers you have set, computed before anything is sent — read it. Use the shortest text you can and the cheapest model; you are measuring a queue, not a voice.
1 · The probe
Fires real generations in parallel and records where the wall is. This is the only lab here that can cost more than pennies if you set the numbers high — the estimate below is the whole spend, before you press anything.
2 · What happened
| # | Wave | Status | Latency | Timeline | Detail |
|---|
One hue for every request: the bar's length is the measurement and its position is when it ran. Red means the request failed, and the status column always says so in words — the colour is never the only signal.
3 · Log
What this tests
A 429 mid-render is the most likely first failure on a small plan written, not run, and it is a bad failure: it arrives after you have paid for the requests that succeeded, halfway through a render, with a partly-built video. The number that prevents it — the semaphore width — is a plan-specific fact that nobody has measured for this account.
The ramp mode fires waves of 1, 2, 4, 8… and stops at the first wave with a failure. The wave before that is your answer.
How to read the result
- The timeline column places each request by when it started and sizes it by how long it took. In a healthy wave the bars start together and end together; when they start together and finish in a staircase, requests are being queued somewhere between you and the model.
- A 429 is not a bug, it is the platform's own bound doing its job — the same class of event as the 402 on the sibling provider, which is the one piece of evidence on this site that a bound is enforceable in practice verified 3 Sep 2026.
- Distinguish 429 from 5xx and from a network error. The vendor's error table maps them: 401 key, 402 quota or plan, 422 validation, 429 concurrency or rate, 5xx theirs vendor docs 5 Sep 2026. A
networkrow in this table usually means CORS or a host CSP, not the API verified 5 Sep 2026.
What to do with the number
Put it in the semaphore. In the pipeline this came from, that is limits.maxConcurrent in the vault's terms file — which is also exactly the field the pattern-three host would enforce on behalf of an app that cannot be trusted to enforce it itself specified, not shipped. Until somebody runs this, that field holds a guess.