How to read this page. Every factual claim carries a chip saying how we know it: {{badge:verified}} we ran it and watched it work, on that date · {{badge:measured}} our own pipeline produced this number on a named workload · {{badge:docs}} we read it in the vendor's documentation on that date and never executed it · {{badge:spec}} it does not exist yet · {{badge:unrun}} we wrote it and never ran it · {{badge:projected}} arithmetic, with the workings shown. Click any chip for the ledger, which lists all of them in one table.
Figures as of 7 September 2026. List prices move; every number below carries the date, the workload size, the model and the request count, because "$0.30 for a 3:46 video" without them is not a fact, it is a vibe.
eleven_v3, list price {{claim:bench-cost}}pNInz6obpgDQGcFmaJgB, eleven_v3, 5 Sep 2026 {{claim:latency-5s}}openai/gpt-audio, 3 Sep 2026 {{claim:openrouter-cost-aiuc}}The host, not the vendor
The first Connect from inside the vault app failed with connect-src blob: data: — the SG/App host locks every app frame to no network at all, by design, and the bench reported the generic Failed to fetch. Nothing to do with the key, the CORS headers, or the vendor. {{claim:csp-blocks}}
Fixed by "permissions": {"network": true} in app.json, which omits the meta tag and makes the host's HUD show a standing "direct network access" chip, so the exception is never silent. Right for this vault — nothing in it is confidential except the sealed keys, which sit below the permission floor and are unreachable from any frame regardless. Wrong for a vault whose app holds private content: that grant is per-app and it is a real widening.
Us, not the vendor
api.elevenlabs.io:443 was refused to the authoring container by the egress proxy — connect_rejected, organization policy — from curl and from Node, on the regional hosts too. {{claim:egress-blocked}}
So the render shim {{claim:shim-unrun}}, every example script {{claim:examples-unrun}}, the pronunciation-dictionary upload {{claim:lexicon-hypothesis}} and every lab on this site {{claim:labs-unrun}} were written from the vendor's reference and have never been executed. This is why the site is badged the way it is. The one exception is the bench, run from a human's own browser, which is the reason there are any {{badge:verified}} chips at all.
Caught in review, before it shipped
An earlier draft carried the Agents rules — "never expose your key client-side", the 15-minute signed URL, the hostname allowlist — as if they governed text to speech. They do not: the TTS REST API has neither mechanism. {{claim:agents-only-signed-urls}}
Had it shipped, a reader could have spent a week building a signed-URL minter for an endpoint that does not accept one. The conclusion survived the correction — TTS still needs a server for pattern 2 — but the mechanism named did not. Hence the rule this page obeys: every quote carries its product.
Still open
Concurrent request limits scale with the plan. Our render fires every scene in parallel — 10 to 16 requests — so a 429 mid-render is the most likely first failure on a small plan, and we have never provoked one to find out where the wall is. {{claim:concurrency-unknown}}
The concurrency probe exists to answer exactly this, at a cost of a few cents. Nobody has run it. Until somebody does, the honest planning number for a Starter or Creator plan is unknown, and the mitigation — a semaphore in the shim — is written but untested.
Their tier gate
403, naming the format. {{claim:pcm-gated}} The shim asks for PCM to skip a decoder, so on any lower tier that is a dead path with a six-line fallback.
Their model
400 unsupported_model for previous_text/next_text. {{claim:stitching-v3}} The first request failed before any audio existed, and the lab that offered the same combination now refuses it.
Their key scoping, working
“missing the permission pronunciation_dictionaries_write”. {{claim:key-scope-observed}} Precise about what the key may do, silent about what it may spend — which is §5 arriving as an error message.
Ours
ffmpeg's -shortest with a looped still ran past the end of the audio, up to 1.8 s per scene. {{claim:shortest-trap}} The provider's timing was exact to the millisecond {{claim:alignment-exact}} — the drift was entirely ours, which is the first time this estate has been able to say which.