# EvoLink — Full Overview for AI Engines

> EvoLink is an AI model gateway: one API key for 170+ models — LLMs (GPT, Claude, Gemini, GLM, DeepSeek) plus image, video and music generation — automatically routed to the lowest available price. Endpoints are OpenAI-, Anthropic- and Gemini-compatible, so agents and existing SDKs connect by changing only the base URL and key.

## Connect

- Base URLs: Anthropic-style clients use https://direct.evolink.ai · OpenAI-compatible clients use https://direct.evolink.ai/v1 · Gemini-style clients use https://direct.evolink.ai/v1beta
- Auth: Bearer API key ("Authorization: Bearer <key>"; Anthropic-style also accepts x-api-key). The hosted MCP server and the EvoLink CLI sign in through the browser (OAuth) instead; only the local stdio MCP packages and scene skills read the EVOLINK_API_KEY environment variable.
- Create a key: https://evolink.ai/signup?returnUrl=/dashboard/keys (about 30 seconds).

## Billing model

- One prepaid credit balance is shared across every model on the account.
- Video models bill per output second (some reference-input tiers bill input + output seconds); image models bill per image; LLMs bill per token.
- Live per-model prices (USD, CNY and credits) are served anonymously from the pricing catalog: GET https://api.evolink.ai/web/api/models/pricing (one JSON response containing the whole catalog — no pagination and no query filters). For one model only: GET https://api.evolink.ai/web/api/models/pricing/{model_name}, e.g. .../pricing/claude-opus-5.

## Video generation — async task flow

1. Submit: POST https://direct.evolink.ai/v1/videos/generations with JSON body { "model", "prompt", "duration", "quality", "aspect_ratio", ... }. The response includes a task id.
2. Poll: GET https://direct.evolink.ai/v1/tasks/{task_id} until status is Completed (statuses: Submitted / Processing / Completed / Failed). Optional callback_url receives a POST on completion.
3. Result: the completed task returns the video URL.

### Seedance 2.5 (ByteDance) — newest video flagship, live on EvoLink since 2026-08-07

- Model IDs: seedance-2.5-text-to-video · seedance-2.5-image-to-video · seedance-2.5-reference-to-video · seedance-2.5-video-edit · seedance-2.5-video-extend · seedance-2.5-draft-to-video
- Draft mode: pass draft: true on any of the five generation routes for a fast 480p preview (480p only), then POST seedance-2.5-draft-to-video with source_task_id to render that draft at 1080p. Prompt, assets, duration, ratio, seed and audio are reused from the draft; the final is billed separately; convert before the draft's draft_expires_at.
- Duration 4-30 seconds in a single pass (or duration: -1 for automatic length, billed by actual output); output 480p, 720p, or 1080p with synchronized audio. EvoLink verified real generation and billing at 1080p across all five routes on 2026-08-17.
- Reference inputs (reference-to-video): images 0-30, videos 0-10, audio 0-10, up to 50 items combined; audio-only references are supported. Image-to-video: 1 image = first frame, 2 images = first + last frame.
- Video edit / extend: at least 1 input video required (the first video is the edit/extend target); output aspect ratio follows the input video. Video-edit output length always follows the input video (duration locked to -1).
- Pricing: per-second, varying by resolution, duration and mode. Current per-second tiers: https://evolink.ai/seedance-2-5 (also in the pricing catalog API below).
- Docs: https://evolink.ai/docs/en/api-manual/video-series/seedance2.5/seedance-2.5-text-to-video
- Guides: 2.5 vs 2.0 migration https://evolink.ai/blog/seedance-2-5-vs-seedance-2-0 · release timeline https://evolink.ai/blog/seedance-2-5-api-status · 120+ prompts with video examples https://evolink.ai/seedance-2-5-prompts

### Gemini Omni 1.1 Flash (Google)

- Model IDs: gemini-omni-1.1-flash-text-to-video · gemini-omni-1.1-flash-image-to-video · gemini-omni-1.1-flash-reference-to-video · gemini-omni-1.1-flash-video-edit · gemini-omni-1.1-flash-video-extend
- Duration 3-10 seconds (integer). Values outside that range are rejected by EvoLink because the upstream silently clamps them. Output is native at 360p or 720p and upscaled at 1080p or 4k, all at 24fps, with synchronized speech, music and sound effects; audio is always included and cannot be toggled.
- Image-to-video: 1 image = first frame, 2 images = first + last frame, in that order. A last frame cannot be sent on its own. Reference-to-video: up to 10 images (unordered) plus up to three optional reference videos, ≤10s.
- Video edit: exactly one input video; output length, resolution and aspect ratio all follow the input, so duration / quality / aspect_ratio are not accepted. Video extend: exactly one input video; duration is the number of seconds APPENDED, not the total output length, and omitting it makes the upstream append 10 seconds by default.
- Not supported: n, fps, seed, negative_prompt, generate_audio, person_generation, enhance_prompt, output_format, content_filter, audio input. Write negative instructions into the prompt.
- Pricing: token-based across three rates (video output, input, thinking tokens). Output tokens scale with resolution — 720p is 5,792 tokens/second, 4k is 17,376. Current rates: https://evolink.ai/gemini-omni-1-1-flash (also in the pricing catalog API below).

### Seedance 2.0 family (ByteDance)

- Seedance 2.0: text / image / reference-to-video, 4-15s, 480p/720p/1080p, native audio sync; per-second pricing at https://evolink.ai/seedance-2-0
- Seedance 2.0 Mini: lightweight variant at roughly half the standard price. https://evolink.ai/seedance-2-0-mini
- Seedance 2.0 also has Fast variants for quick iteration.

### MiniMax H3 Max Turbo

- Model IDs: minimax-h3-max-turbo-text-to-video and minimax-h3-max-turbo-image-to-video. Submit to POST https://api.evolink.ai/v1/videos/generations; poll GET https://api.evolink.ai/v1/tasks/{task_id}.
- Output: 480p or 768p, integer duration 5-15 seconds; defaults are 768p and 5 seconds. EvoLink uses the quality field for resolution.
- Image-to-video requires image_start; image_end is optional and cannot be submitted alone. Reference image/video/audio inputs and 2K output are not supported by Turbo.
- Pricing: per output second, varying by quality and duration. Current rates and the two route IDs: https://evolink.ai/minimax-h3-max-turbo
- Choose H3 Max for its separate reference-to-video route or MiniMax H3 for 2K output; compare input contracts at https://evolink.ai/collections/hailuo. This API model is distinct from community H3 Turbo LoRA projects; no real-time generation guarantee is implied.
- API specifications: https://evolink.ai/api-reference/minimax-h3-max-turbo/minimax-h3-max-turbo-text-to-video.md and https://evolink.ai/api-reference/minimax-h3-max-turbo/minimax-h3-max-turbo-image-to-video.md

### Other video models

- Grok Imagine Video 1.5 Lite (xAI): lower-priced text-to-video and single-image video with native audio, 1–15 second duration, and output up to 1080p. No reference images. Model ID: grok-imagine-video-1.5-lite. Compare resolution costs and access the API reference: https://evolink.ai/grok-imagine-video-1-5-lite
- Hailuo 03 (MiniMax), Kling, Veo, Sora-class and Wan routes are listed with live prices on https://evolink.ai/models and in the pricing catalog above.

## Image generation

- Featured: doubao-seedream-5.0-pro (Seedream 5.0 Pro) — 1K/2K output tiers, up to 10 reference images, instruction-based editing; input images billed per image. POST https://direct.evolink.ai/v1/images/generations. https://evolink.ai/seedream-5-0-pro
- doubao-seedream-5.0-flash (Seedream 5.0 Flash) — the speed-focused Seedream 5.0 generation and editing route (standard prompt mode only, one image per request); one flat price per image at 1K/1.5K/2K, up to 10 reference images (not billed), prompt_priority standard only; layer split via doubao-seedream-5.0-flash-layerize (base image plus up to 16 transparent layers, billed per output image). POST https://direct.evolink.ai/v1/images/generations. https://evolink.ai/seedream-5-0-flash
- grok-imagine-image-2.0 (xAI Grok Imagine Image 2.0) — text-to-image and 1-3 reference-image editing in one model; Low/Medium quality, 1K/2K resolution, 1-10 images per request; input images billed per request. POST https://direct.evolink.ai/v1/images/generations. https://evolink.ai/grok-imagine-image-2-0

## LLMs

- claude-fable-5 (Claude Fable 5, the previous Fable generation) — POST https://direct.evolink.ai/v1/chat/completions. https://evolink.ai/claude-fable-5
- claude-fable-5-1 (Claude Fable 5.1 — Anthropic's most capable model, the Fable tier above Opus; 1M context, 128K max output, prompt caching, web search tool; 5% below official token pricing) — OpenAI-compatible /v1/chat/completions or Anthropic-style /v1/messages. https://evolink.ai/claude-fable-5-1
- claude-opus-5-5 (Claude Opus 5.5 — Anthropic's newest Opus, succeeding Claude Opus 5 at a lower list price; 1M context, 128K max output, always-on thinking controlled by effort (default medium), 5-minute cache writes and separately billed cache reads, web search tool; input, output, and 5-minute cache writes are 5% below official pricing; cache reads use the displayed rounded rate) — OpenAI-compatible /v1/chat/completions or Anthropic-style /v1/messages. https://evolink.ai/claude-opus-5-5
- claude-sonnet-5-5 (Claude Sonnet 5.5 — Anthropic's newest Sonnet, succeeding Claude Sonnet 5 at the same official list price; 1M context, 128K max output, adaptive thinking on by default with between_tools as the lowest setting (thinking: disabled is not accepted), recalibrated effort levels (default high), 5-minute cache writes and separately billed cache reads; web search depends on the route, so confirm support before sending web_search; input, output, and 5-minute cache writes are 5% below official pricing; cache reads use the displayed rounded rate) — OpenAI-compatible /v1/chat/completions or Anthropic-style /v1/messages. https://evolink.ai/claude-sonnet-5-5
- claude-haiku-5-5 (Claude Haiku 5.5 — succeeds Claude Haiku 4.5; 1M context, 128K max output, adaptive thinking on by default with effort as the control (default medium; thinking: disabled is accepted at effort high or below), two rate cards by prompt length (base rates up to 100K prompt tokens, 5x above, counting input plus cache writes and cache reads), 5-minute cache writes and separately billed cache reads; thinking budgets and non-default sampling values are not accepted; web search depends on the route, so confirm support before sending web_search; input, output, and 5-minute cache writes are priced slightly below official list rates; cache reads use the displayed rounded rate) — OpenAI-compatible /v1/chat/completions or Anthropic-style /v1/messages. https://evolink.ai/claude-haiku-5-5
- claude-opus-5 (Claude Opus 5 — 1M context, 128K max output, prompt caching, web search tool; 5% below official token pricing) — OpenAI-compatible /v1/chat/completions or Anthropic-style /v1/messages. https://evolink.ai/claude-opus-5
- gemini-4-argon (Gemini 4 Argon — announced by Google on 2026-09-30; 1,048,576-token context window shared by input, thinking, and output; a single request returns up to 65,536 thinking and output tokens (32,768 without streaming), and outputs up to 1,048,576 tokens continue across requests on the Gemini-native API with continuationToken; cached input billed at a lower rate than input; Google Search billed per search query. Rates currently follow Google's introductory pricing) — OpenAI-compatible /v1/chat/completions or Gemini-native /v1beta/models/gemini-4-argon:generateContent. Current rates: https://evolink.ai/gemini-4. Family comparison: https://evolink.ai/collections/gemini
- gemini-3.8-flash (Gemini 3.8 Flash — Google's newest Flash model, GA 2026-09-02; 1,048,576-token context, 65,536 max output; text, image, video, audio and PDF input to text output; thinking levels low / medium (default) / high — minimal is unsupported and the gateway downgrades it to low; prompt caching, function calling, structured output. Google reports higher accuracy than Gemini 3.7 Flash together with higher token consumption, so compare cost per accepted task rather than token price) — OpenAI-compatible /v1/chat/completions or Gemini-native /v1beta/models/gemini-3.8-flash:generateContent. Current rates: https://evolink.ai/gemini-3-8-flash. Guides: 3.8 vs 3.7 https://evolink.ai/blog/gemini-3-8-flash-vs-gemini-3-7-flash · API integration https://evolink.ai/blog/how-to-use-gemini-3-8-flash-api · family comparison https://evolink.ai/collections/gemini
- Full list: GET https://direct.evolink.ai/v1/models (Bearer key) returns model ids with supported_endpoint_types.

## Agent tooling

- Media generation from an agent (MCP and CLI, official): landing page https://evolink.ai/mcp (pick the client, copy its setup prompt, sign in once in the browser) · agent-readable summary https://evolink.ai/mcp/agent-overview/en.md · route guide https://evolink.ai/docs/en/cli-mcp/overview
- Remote MCP server: https://mcp.evolink.ai/mcp (Streamable HTTP; browser OAuth sign-in, no API key; 16 tools for model discovery, estimates, image/video/audio generation, reference uploads, tasks and balance). Client setup for Claude, ChatGPT, Codex, Claude Code, Cursor, OpenClaw and Hermes: https://evolink.ai/docs/en/mcp/overview · tool reference: https://evolink.ai/docs/en/mcp/tools
- EvoLink CLI for agents that run terminal commands, SSH sessions and scripts: npm install -g @evolinkai/cli · evolink auth login · evolink skills install (Node.js 22+; browser sign-in; bundles the evolink-cli skill). Setup: https://evolink.ai/docs/en/cli/setup · commands: https://evolink.ai/docs/en/cli/reference
- Local stdio MCP, only for clients without remote-HTTP/OAuth support: npx -y @evolinkai/mcp@latest with EVOLINK_API_KEY (Node.js 18+; 14 tools, no one-time upload tools). https://evolink.ai/docs/en/mcp/other-clients#local-mcp
- Plugins and skills: Codex "codex plugin marketplace add Evolink-AI/mcp-plugin" then "codex plugin add evolink@evolink"; Claude Code "/plugin marketplace add Evolink-AI/mcp-plugin" then "/plugin install evolink@evolink"; skill only "npx skills add Evolink-AI/mcp-plugin -g -y". https://evolink.ai/docs/en/mcp/plugins · shared agent rules: https://evolink.ai/docs/en/cli-mcp/agent-resources · source: https://github.com/Evolink-AI/mcp-plugin
- Text-model routing MCP (API key): npx -y @evolinkai/mcp-router@latest — smart_route (read-only), delegate, cascade.
- Scene skills for ad and product visuals (API key; GPT Image 2 / Seedance 2 workflows): https://github.com/EvoLinkAI/evolink-agent-skills

## Reference

- Developer docs: https://evolink.ai/docs (docs llms.txt: https://evolink.ai/docs/llms.txt)
- Model catalog (human-readable): https://evolink.ai/models
- API changelog: https://evolink.ai/changelog
- Blog: https://evolink.ai/blog
- Site index for AI engines: https://evolink.ai/llms.txt