docs: add Y-API provider page

[jiweiyeah]

What

Adds the docs page for the Y-API provider: docs/providers/y-api.md, plus the
corresponding sidebars.js entry in its alphabetical slot (between xinference and zai).

Why

Companion to BerriAI/litellm#41559, which registers Y-API as a JSON-configured
OpenAI-compatible provider. That PR's provider_endpoints_support.json entry points its
url field at:

"url": "https://docs.litellm.ai/docs/providers/y-api",

which 404s until this page exists. The alternative — dropping url the way chutes and
poe do — is also fine if you'd rather keep the matrix link-free; I went with the page
since the provider is JSON-configured and therefore has no code module to carry the
documentation.

Page contents

• Overview table — provider route y-api/, base URL, the four supported operations.
• Required variables — YAPI_API_KEY.
• Usage — Python SDK (non-streaming + streaming), proxy server, legacy /completions,
  /responses, and /v1/messages through the proxy.
• Model list — the fifteen models GET /v1/models currently returns, with prices.
• Two-slash model ids — y-api/deepseek/deepseek-v4-flash resolves on the first slash
  and forwards deepseek/deepseek-v4-flash upstream unchanged. Worth calling out because
  every Y-API model id carries an upstream org prefix, so every model string has two
  slashes. There is a test for exactly this case in the main PR.
• Reasoning — the relay accepts a narrower reasoning_effort set than the labs it
  fronts, so the page lists what each model actually takes (gpt-5.6-* up to xhigh,
  gpt-6-astra no none, tencent/hy3 only low/high). All measured against the live
  API, n=5 per level.
• Token limits — the openai/* models require max_completion_tokens.
• Embeddings — not served, so litellm.embedding() is unsupported.

I verified every code sample on this page against the live endpoint through a local
LiteLLM proxy built from the branch in #41559.

Checklist

• docs/providers/y-api.md added
• sidebars.js updated in alphabetical position
• Every example verified end-to-end against the live API
• No url in the main PR's matrix points at a 404 after this lands

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
litellm	Ready	Preview	Oct 4, 2026 7:06am UTC

[jiweiyeah]

Correction pushed in e2968f5402.

The "Token limits" section claimed LiteLLM already sends max_completion_tokens for the openai/* models. That was wrong — the parameter is forwarded verbatim, which is exactly why the provider entry in BerriAI/litellm#41559 now declares a max_tokens → max_completion_tokens mapping. The section now says both spellings behave identically through LiteLLM.

Verified by pointing the provider at a local echo server and reading the outbound body: max_tokens=16 produced {"max_tokens": 16} before the mapping and {"max_completion_tokens": 16} after it.

[jiweiyeah]

It's been ~2.5 weeks since this was opened (Sep 17) and it's been green since: Vercel preview deployed, and the docs page builds. Nothing has changed on my side and the content is one page plus the nav entry.

If the holdup is anything specific — the docs.litellm.ai/providers/advanced/write-pass-through guidance, the base URL quoting style, or anything else — I'll adjust it today. mergeable is clean; the only remaining blocker is a maintainer review.

Also pinging in case it got buried: the sibling provider-registration PR (BerriAI/litellm#41559) and its companion fix (BerriAI/litellm#43784) are still open. Happy to reorder or split them, whichever is easier for you to land.

[jiweiyeah]

Refreshed today after a pricing-basis change on our side (6a4211a9), so the table is correct at merge time:

• The limited-time 1:20 top-up promo ended 2026-10-01; the conversion is now fixed at $1 paid = $10 credit (top_up.quota_rate: 10, is_promotional: false in pricing.json, read 2026-10-04). Every cash figure in the previous version of the table was computed under the old promo, so all of them doubled.
• The table now lists all 20 models the gateway serves today (values read from pricing.json), where the old snapshot had 15 and was missing seven current models. openai/gpt-5.6-terra and openai/gpt-6-astra were dropped: both return 503 "No available channel" and are absent from GET /v1/models as of today.
• The page now states the conversion with its history and cites pricing.json with the snapshot date, so a future rate move is visible rather than silently wrong.