elevenlabs.providers.sgit.ai

ElevenLabs — text to speech with character-level timestamps

This is a report, not a tutorial. The vendor's documentation explains the API better than we can and stays fresher; what follows is what happened when we actually used it — what it cost on a named workload on a named date, what broke, and which credential patterns the product can support. Sections 8 and 9 are the point of the page.

Prose from the video vault at commit 7d1916aca5f3, 7 September 2026. Prose from platforms/elevenlabs.md, with the figures behind it from four reels' worth of pipeline findings. When the vault moves ahead, this page is behind — and says so rather than guessing.

What this site is, and is not

It is narrower than the vendor's own environment, not better. The vendor serves millions of users across every workload; this site serves one — narrating explainer videos from a scripted pipeline — and reports what that cost, where the key had to live, and what broke. A general-purpose interface cannot be narrow, which is why this one does not try to be: the labs here exist because the vendor's own console does not fit this workflow, not because it is inadequate.

It is a report with a workbench attached, not a tutorial and not an index of somebody else's documentation. If a section here could be replaced by a link to the vendor, it should be deleted, and the vendor's page is linked with the date we read it. What this site has that nobody else does is what it cost on a named workload on a named date, what went wrong, and which credential patterns the product can actually support.

"Provider" here means one thing only: a service that serves models over an API — the sense this estate's own code uses, where a constant of that name holds an entry per model service. It does not mean a customer who redistributes capability to the people they serve, which is the other established sense in this estate and is not what any page here is about.

It is not an endorsement, and not the vendor's. No commercial relationship exists, the page saying so shipped before there was anything to disclose, and every claim carries the state that says how we know it.

1 · Disclosure

None. No credits, programme, or commercial relationship with ElevenLabs as of 5 September 2026. The key used for the one session reported in §8 is a Creator-tier key we pay for ourselves at list price. The full list — which currently has no entries in the "relationship" column — is at /disclosures/, and it shipped with this site's first version precisely so that the page's existence is not read as evidence of a relationship.

2 · What it grants

The rows an agent's grant gains when this platform is connected, as capability tuples — verb × object class × reach, with reversibility marked. This is the block a capability index joins on; it is emitted as front-matter in this page's markdown as well as in the table below.

What the platform can grant, per product

What our key grants — scoped to text to speech and voices-read

Not granted by a key scoped this way, and it should stay that way: voice cloning, dubbing, agents.

The same rows are emitted as front-matter in this page’s markdown source, so a capability index can join across providers without parsing English.

Read the two tables as one argument. The first is what an agent gains the moment this provider is connected at all; the second is what the key this evaluation actually used can reach. The gap between them is the work a scope does, and it is the only part of a credential story this vendor lets you control.

The row worth stopping on is create × voice × tenant. Connecting this provider gives an agent the capability to produce a voice that did not exist — instant cloning takes under two minutes of audio. That is a different order of thing from producing audio, no vendor page frames it as a capability, and the platform's controls for it are procedural rather than technical: professional cloning requires verification of the speaker, and instant cloning requires you to have the right to the recording you upload. Consent is not something the API can check. Our key is scoped to exclude it, and the key-scope probe is how you check that yours is too.

One row a reader might expect is deliberately absent. It has been proposed that connecting a voice platform grants send × audio × world, irreversibly, on the reasoning that audio sent is sent. Not for this product. The text-to-speech endpoint returns audio to the caller; nothing leaves for the world, and the estate's own primitive distinguishes endpoint reach from world reach exactly here. The irreversible row this vendor does offer is Agents, in the table above, and it is scoped out of our key. Saying otherwise would tell a reader that generating narration publishes it — which is the error §4 exists to prevent, one product over. vendor docs 5 Sep 2026

Two more things are worth saying in prose. First, read × audio+timing × self is the new capability — this API returns the start and end time of every character it speaks, which the pipeline that drove this evaluation never had from any other provider. Everything in the captions studio is arithmetic on that array. Second, nothing our key grants reaches the world. The pipeline receives audio; publishing a video is a separate, human act by a person who watched it first. That distinction is the difference between a narration tool and a voice agent, and it is why Agents is scoped out of the key rather than merely unused.

3 · Which pattern — and for which product

A vendor with several products has several answers, so this section is indexed by product, not by vendor. The four patterns are defined on /patterns/; the generated matrix, including the sibling provider, is at /comparison/.

Text to speech (the REST API — what we use).

• Pattern 0, key in the page: technically available. CORS permits a browser to call api.elevenlabs.io directly verified 5 Sep 2026 — and the only bound on that key is the account's monthly quota vendor docs 5 Sep 2026. So a key in a published page is pattern 0 with a ceiling. Never publish one. Every lab on this site is the narrower case: the key's owner, testing their own key, in their own browser.
• Pattern 1, bounded key in the page: not available. ElevenLabs keys are scoped, not bounded — see §5.
• Pattern 2, short-lived token: available only with a server we run. The vendor offers no short-lived credential for TTS vendor docs 5 Sep 2026. Such a minter would be ours to build, host and pay for; it does not exist.
• Pattern 3, host holds the key: specified, not shipped specified, not shipped.

Agents (conversational). Pattern 2 with the vendor's own signed URL (15 minutes), or a hostname allowlist — and the vendor's documentation says not to configure both on one agent vendor docs 5 Sep 2026. These mechanisms are the reason §4 exists, and they do not apply to text to speech.

4 · Where the key goes

Every quote below carries the product it belongs to, the URL, and the date we read it. This is not pedantry: an earlier draft of our own strategy quoted the Agents rule on a TTS page, which would have sent a reader off to build a signed-URL minter for an endpoint that does not accept one.

Agents product — read 5 September 2026 at https://elevenlabs.io/docs/eleven-agents/customization/authentication: "Never expose your ElevenLabs API key client-side." Signed URLs "are valid for 15 minutes… the conversation must be initiated within the 15 minute window"; allowlists use exact hostname matching; "Do not configure signed URLs and allowlists together on the same agent." vendor docs 5 Sep 2026

Text-to-speech product — there is no equivalent page. The key goes in the xi-api-key request header vendor docs 5 Sep 2026, and the only vendor-side control is scoping a key to endpoints in the dashboard. No signed URL, no allowlist, no origin check. A TTS key is a bearer credential for the account's quota, and the vendor's client-side warning above is written about a different product — which does not make it wrong here, only unenforceable here.

Where the key goes in our own estate, for completeness: owner-sealed at .vault/elevenlabs/config.json (AES-256-GCM under a key derived from the vault's write key, so a read-key holder cannot open it), opened for the life of one shell command. A key pasted into a lab on this site goes to this browser's localStorage and nowhere else verified 7 Sep 2026.

5 · The bounding primitive

The plan's monthly character quota, reported by GET /v1/user/subscription vendor docs 5 Sep 2026. That is the whole bound, and it belongs to the account, not to the key.

Per key, ElevenLabs gives you scope, not spend: a key can be restricted to endpoints — text-to-speech yes, dubbing no — which limits what a leaked key can do, and does nothing at all to limit how much of your quota it can burn vendor docs 5 Sep 2026.

What it does not cap:

• Any single key can consume the whole account's quota.
• Nothing resets faster than the billing cycle — there is no daily or weekly window to fall back to.
• There is no per-key limit, and therefore no equivalent of the thing that saved us on the sibling provider: a refusal at a ceiling you chose in advance verified 3 Sep 2026.

A leaked ElevenLabs key is bounded by the plan. A leaked OpenRouter provisioned key is bounded by the number you typed when you minted it. That difference is the entire reason the comparison table earns its place on this site.

6 · The minimal working example

The smallest thing that runs, as a file rather than a snippet — and every one of these is written from the vendor's reference and never executed written, not run, because the container that wrote them could not reach the API verified 4–5 Sep 2026.

The pipeline invocation they were written for:

export ELEVENLABS_API_KEY=$(node tooling/scripts/vault-secrets.mjs open elevenlabs)
node examples/tts-timestamps.mjs "<text>" <voice_id> /tmp/out   # → .wav + .words.json + .srt

The interactive version is the bench, which has run: it is the source of every verified chip on this page. The rest of the labs — twelve of them — extend the bench to the features this evaluation did not reach, and they are all written, not run.

7 · What we use it for

Nothing in a shipped video yet, and it is worth saying that plainly before the numbers.

The workload is a video pipeline that renders explainer reels from a reel.json script: one narration string per scene, rendered to audio, composited over stills with the narration text on screen as a caption band. It has run with three speech providers — a local Kokoro model ($0), an OpenRouter-hosted gpt-audio (paid, measured in §8), and now this one, evaluated. Four reels, six cuts, are published.

What this API would change, ranked by how much it improves those videos rather than by how impressive it is:

1. Character-level timestamps — SRT sidecars, chapters accurate to the sentence, and a word-synced caption band. None of it is possible with either incumbent provider, and all of it is arithmetic on one response field. This is the reason for the evaluation.
2. A speed control back. voice_settings.speed, 0.7–1.2 vendor docs 5 Sep 2026. The incumbent has none, and the cost of that was a portrait cut losing a third of its script rather than 15% of its pace measured 2 Sep 2026.
3. A pronunciation dictionary, so the narration text can say sgit.ai instead of sgit dot ai — because in this pipeline the narration text is the caption text, so every pronunciation hack is on screen where viewers read it as a typo vendor docs 5 Sep 2026.
4. Forced alignment of the six existing cuts — subtitles for videos already published, including the two Kokoro cuts that will never have API timestamps, for about two cents vendor docs 5 Sep 2026.
5. Scribe as a QA gate — transcribe the finished file, diff it against the script, and catch automatically the two failures we have shipped and then caught by eye.

Items 1–3 have a lab each on this site. None of the five has been run.

8 · What it cost

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.

What we actually paid ElevenLabs, in full: one bench session on 5 September 2026 — four generations of the same 187-character sample on eleven_v3 with timestamps, about 750 characters, $0.075 at the list rate of $0.10 per 1,000 characters vendor docs 5 Sep 2026. That is the entire measured spend. Everything else in this section is arithmetic, and is labelled as such.

Projected, with the arithmetic shown. Our scripts run about 6 characters per word; the word counts are measured from the four reels we have rendered measured 3 Sep 2026. Characters × the model's list rate:

projected for the two ElevenLabs columns — no reel has been rendered with this API. measured for the words and for the right-hand column. The conclusion the arithmetic supports: v3 costs about what the incumbent costs; flash costs half — and both are a rounding error against the value of the timestamps, which is the only sound reason to switch.

And one more measured figure, added 8 September 2026: the first video made with this API — nine scenes, 1,821 characters, 2:08 of speech, two cuts — cost $0.18 at list price and was paid in a free tier's quota rather than in money verified 8 Sep 2026. Zero re-renders, because the script was written first. The full account is on /video/.

Plan arithmetic. At about 4,000 characters per video, a Creator plan's 220,000 monthly characters is roughly 50 videos a month projected. Rates and quotas: vendor docs 5 Sep 2026, vendor docs 5 Sep 2026.

Run the arithmetic on your own script in the cost model — it needs no key and makes no network call.

9 · What went wrong

The section nobody else writes. Four things, in the order they cost us time.

And what went wrong the first time somebody ran it

On 8 September 2026 the render path ran for the first time verified 8 Sep 2026. Three of the four things that broke are the vendor's, and one is ours:

What has not gone wrong, because nobody has tried it

The most useful thing in this section is the list of failures we have not yet had the opportunity to have. written, not run written, not run written, not run Every one of them has a lab on this site, and each lab prints a result you can paste back into the vault as evidence.

The full list of what remains unverified — and who could verify it — is the open items, which is the last section of the ledger.

Every claim on this site, with its state → Disclosures →