Kunavo

Kunavo resells partner models (Gemini, Claude, …) at a discount — roughly 20% off the vendor's own price. It's the one provider that spans both worlds: text (an OpenAI-compatible endpoint) and media (image + video, via an adapter).

Text

import { createOpenAICompatible } from "@ai-sdk/openai-compatible";

const kunavo = createOpenAICompatible({
  name: "kunavo",
  baseURL: "https://api.kunavo.com/v1",
  apiKey: process.env.KUNAVO_API_KEY,
});
// dashed ids: kunavo("gemini-2.5-flash"), kunavo("claude-sonnet-4-6")

Model ids are dashed (gemini-2.5-flash, claude-sonnet-4-6).

A model vendor's own official API (@ai-sdk/openai, @ai-sdk/anthropic, …) is just another entry in the list — often the cheapest, and least likely to break native features like prompt caching or tool calls.

Image & video

For media, wire Kunavo with createKunavoMediaAdapter and use it in createMediaLCR. All paths below were verified live against the real API (image 2026-05-31, edit + async/sync video 2026-06-06).

import { createKunavoMediaAdapter } from "ai-lcr";

const kunavo = createKunavoMediaAdapter({
  apiKey: process.env.KUNAVO_API_KEY!,
  videoMode: "async",       // default; "sync" uses the blocking single-call path
  pollIntervalMs: 5000,     // async-video poll cadence (default)
  pollTimeoutMs: 600_000,   // give up on an async job after this (default 10m)
});

Image — generate and edit

Text-to-image hits POST /v1/images/generations. Routes whose externalId ends in -edit (nano-banana-edit, gpt-image-2-edit) are reference-image edits and route to POST /v1/images/edits instead — the caller supplies image (url or data-uri) or image_urls[] in the input.

registry: {
  "google/nano-banana-pro": {
    id: "google/nano-banana-pro",
    modality: "image",
    routes: [
      { provider: "kunavo", externalId: "nano-banana-pro", pricing: { unit: "image", cents: 6.7 } },
    ],
  },
}

Video — Veo, async submit→poll (sync fallback)

Kunavo serves Google's Veo family: veo-3, veo-3-lite, veo-3-quality. By default the adapter uses the async path — POST /v1/videos to submit, then polls GET /v1/videos/{id} (queued → in_progress → completed) and returns the clip URL. This is Kunavo's recommended production path: non-blocking, and it survives proxy/load-balancer connection limits on long renders. A poll timeout fails over to the next route.

Set videoMode: "sync" to use the blocking single call POST /v1/video/generations instead (capped by syncVideoTimeoutMs, default 10m).

registry: {
  "google/veo-3-lite": {
    id: "google/veo-3-lite",
    modality: "video",
    routes: [
      // flat per-clip price; verify clip duration/resolution/audio before trusting it
      { provider: "kunavo", externalId: "veo-3-lite", pricing: { unit: "call", cents: 16 } },
    ],
  },
}

Kunavo video ids are dashed and billed per call (flat per clip), not per second — so on a short clip they can beat per-second providers by a wide margin. But GET /v1/models exposes no pricing, so the per-call cents are hand-entered: confirm clip duration, resolution, and audio against your own render before trusting the gap.

See Image & video for the full media flow and the bundled model catalog.