elevenlabs.providers.sgit.ai / experiments / Cost model — no key, no network
Cost model — no key, no network
Arithmetic you can check, on text you paste, with no key and no request to anybody. Cost figures go stale faster than anything else on this site — this page makes the staleness visible instead of hiding it in a table.
No key bar on this page, because there is nothing to send. Everything below is computed in your browser from the text in the box and the list prices hard-coded in lab.js. It is the only lab here that cannot cost you anything.
This lab needs no key and makes no network call at all. Everything below is arithmetic in your browser on text you paste. It is here because a cost model you can check is worth more than a price you are told.
1 · Your script
2 · What it would cost
| Model | Rate / 1k chars | Characters | Requests | Cost | Per-request limit |
|---|
Rates are the vendor's September 2026 list prices as read on 5 September 2026, hard-coded in lab.js where you can check them. They move. Nothing here calls the vendor to confirm them.
3 · Against what we actually paid
| Reel | Words | Chars | Paid (gpt-audio, measured) | flash here | v3 here |
|---|
The left-hand numbers are measured invoices from a different provider on named dates; the two right-hand columns are this page's arithmetic. Only one of those is evidence.
4 · Does a plan cover it?
The arithmetic, stated
Cost = characters ÷ 1,000 × the model's rate. Rates, read 5 September 2026: eleven_v3 and eleven_multilingual_v2 $0.10 per 1,000 characters; eleven_flash_v2_5 and eleven_turbo_v2_5 $0.05 vendor docs 5 Sep 2026. Context text passed as previous_text or next_text is not billed as generated characters, which is why stitching is free.
Two things this page counts that a rate card does not:
- Requests, not just characters. A scene-by-scene render fires one request per scene, and the request count is what meets the concurrency limit — the failure mode that costs a render, rather than a rounding error on the bill written, not run.
- Your longest scene against the model's per-request limit.
eleven_v3takes 5,000 characters,multilingual_v210,000, flash and turbo 40,000 vendor docs 5 Sep 2026. Our longest scene is about 420 characters, so limits are not a constraint for a scene-by-scene render — they bite only if somebody concatenates a whole reel into one request, which also destroys the per-scene timing the pipeline needs.
Against what we actually paid
Section 3 puts this page's arithmetic beside real invoices from a different provider on named dates measured 3 Sep 2026. Only the left-hand column is evidence; the two right-hand columns are this page doing multiplication projected.
The conclusion that survives the comparison: v3 costs about what the incumbent cost, flash costs half, and both are a rounding error against the value of the timestamps. The reason to switch is not price.
Plan fit
At about 4,000 characters per video, a Creator plan's 220,000 characters is roughly 50 videos a month projected — arithmetic, not experience; we have never run a month of it. Section 4 does the same division on your own numbers. What it cannot tell you is whether the plan's concurrency allowance survives your render, which is the other half of choosing a tier and is a different lab.
When these numbers go wrong
The day the vendor changes a price. Everything here carries the date it was read, and the ledger marks the whole class of pricing claims as vendor-documentation rather than measurement vendor docs 5 Sep 2026. If you find a rate that has moved, that is a correction worth sending: it invalidates a row on this site rather than a paragraph.