title: OpenRouter — the sibling page, as a structural stub description: "Not this site's subject. A stub of the sibling provider page, carried here to prove that adding a provider is one Markdown file with front-matter — and to supply the comparison table's other rows." lead: "This is not an OpenRouter site. It is a deliberately short stub of the sibling provider page, published here for two reasons: it supplies the other rows of the comparison↗, and it is the proof that adding a provider to this family is one Markdown file with front-matter rather than template surgery." order: 70 provenance: commit: 7d1916aca5f3 date: 7 September 2026 note: "Condensed from platforms/openrouter.md. The full page belongs at openrouter.providers.sgit.ai, which does not exist yet." grants:

• verb: spend object: budget reach: tenant reversible: true bounded_by: the key's own limit, with a monthly reset — a number chosen in advance
• verb: send object: text reach: endpoint reversible: true bounded_by: each scene's narration goes to the model
• verb: read object: audio reach: self reversible: true bounded_by: "the speech comes back; the platform publishes nothing" not_granted: [publishing, world-visible side effects] patterns:
• provider: OpenRouter product: Chat and audio models server: "No" p0: verdict: never note: "possible; never acceptable" p1: verdict: yes note: "provisioned key with a limit and a reset" p2: verdict: yes note: "with a server, and not needed" p3: verdict: yes note: "shipped: sg.llm.* in the SG/Vault host"

  ---

<div class="note"><p><b>Scope.</b> This site reports on ElevenLabs. The full OpenRouter report — nine sections, its own cost table, its own failures — belongs at <code>openrouter.providers.sgit.ai</code> when that site exists. What follows is the minimum needed to make the comparison honest, plus the one measurement that is genuinely evidence for the whole argument.</p></div>

Why it is the useful contrast

OpenRouter can mint a key with a spend limit and a reset window, and reports limit, usage and limit_remaining against it. So a key in a page has a blast radius somebody chose in advance. ElevenLabs text to speech has no equivalent at any price {{claim:no-per-key-spend-limit}} — which is the single sharpest difference between the two and the reason the comparison table earns its place.

What it cost us, measured

About $0.08 per minute of speech {{claim:openrouter-cost-aiuc}}. Speech generation took 3.3–5.4 s for a whole reel in parallel, against 50–94 s for a local model. Transcript mismatches across 61 generations: zero {{claim:transcript-mismatches}}.

What went wrong

• The limit did its job — a 402 at $4.79 of $5.00, because audio output requires $0.50 of headroom. The full framing is on the patterns page↗, where it belongs: it is the one piece of evidence on this site that a bound is enforceable in practice rather than in a diagram {{claim:openrouter-402}}.
• No speed control. The model takes no pace parameter, so a portrait cut that had to fit under a three-minute limit lost a third of its script instead of 15% of its pace {{claim:openrouter-no-speed}}. That single missing knob is what sent us to evaluate a vendor whose API has one {{claim:speed-range}}.
• An outro that lied. A reel's closing narration said "no API cost" over a slide printing $0.1554, because one outro served both a free cut and a paid one. Caught by reading the closing frame; the rule now is that every outro says "the cost is on the screen".

This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).