API Updates

EvoLink MCP and EvoLink CLI are live. Give the agent you already use access to EvoLink's 150+ image, video and audio models: Claude, ChatGPT, Codex, Claude Code, Cursor, OpenClaw, Hermes and other MCP clients. Sign in once in the browser; there is no API key to paste.

New

• Remote MCP server: https://mcp.evolink.ai/mcp (Streamable HTTP). On first use the client opens a browser to sign in to EvoLink and approve the connection. 16 tools: find and compare models, read parameters and prices, estimate cost, generate images, video and audio, upload references, check tasks and balance
• EvoLink CLI: for agents that run terminal commands, SSH sessions and scripts. Install with npm install -g @evolinkai/cli, sign in with evolink auth login, then evolink skills install adds the companion skill for Codex, Claude Code, Cursor, OpenClaw, Hermes, Gemini CLI, OpenCode and Copilot CLI. Requires Node.js 22+
• MCP page: pick your client on evolink.ai/mcp, copy the setup prompt and send it to your agent. An agent-readable summary is available at /mcp/agent-overview/en.md
• Plugins and local MCP: Codex and Claude Code can install the EvoLink plugin from the Evolink-AI/mcp-plugin marketplace. Clients without remote MCP support can still run the local stdio package @evolinkai/mcp with an API key
• Docs: a new CLI / MCP section with a route guide, per-client setup, tool and command references, and troubleshooting

Notes

• Billing: MCP and CLI are free, and lookups cost nothing. Generation is charged to your EvoLink balance at the model's price, the same as the API. Ask your agent for an estimate before approving a paid task
• References: pass public links, or upload local files; a one-time upload URL accepts files up to 95 MB, and the CLI uploads directly
• Access: each connection is a separate authorization. Disconnect it in your client's MCP settings, or revoke it from your EvoLink account
• Open the MCP page | CLI / MCP docs

Claude Haiku 5.5 from Anthropic is now live on EvoLink — the successor to Claude Haiku 4.5 for high-volume classification, extraction, routing, and sub-agent work, with a 1M-token context window and up to 128K max output.

Claude Haiku 5.5

• Model ID: claude-haiku-5-5
• Context window: 1M tokens
• Max output: 128K tokens
• Capabilities: tool use, structured output, adaptive thinking (on by default) with effort control, vision input, prompt caching
• Recommended workloads: high-volume classification and extraction, request routing, and sub-agent tasks
• Billing: token-based — Input, Output, Cache Write, Cache Hit, on two rate cards selected by prompt length (up to 100K prompt tokens, and above); the web search tool is billed per search on routes that support it. See current rates
• API: Anthropic Messages, with the same EvoLink API key as your other models
• Try Claude Haiku 5.5

HappyHorse, Wan and Qwen Image 3.0 get lower prices on EvoLink, effective October 8, 2026 (UTC+8). The update covers HappyHorse 1.0, HappyHorse 1.1, Wan 2.6, Wan 2.6 Flash, Wan 2.7, Qwen Image 3.0 and Qwen Image 3.0 Pro: prices are lower at every resolution and on every billed item of these models. The new rates apply automatically, with no changes to your API calls.

• Wan 3.0 and Wan 3.0 Prime prices are unchanged
• Qwen3.8 Max pricing has been adjusted
• HappyHorse 1.0 | HappyHorse 1.1 | Wan 2.6 | Wan 2.7 | Qwen Image 3.0 | Qwen3.8 Max | Pricing

Qwen Audio 3.1 TTS Flash is now on EvoLink's unified audio API: text-to-speech with 68 system voices, plus a companion model, Voice Enrollment, that creates your own voice for it.

New

• Text-to-speech: model ID qwen-audio-3.1-tts-flash. Send up to 5,000 characters of text and get MP3, WAV or Opus audio back. Control emotion, pace and dialect with a short instruction, or with inline tags such as [excited]
• Voice cloning and voice design: model ID voice-enrollment. Send audio_url to clone a voice from a short recording, or voice_prompt and preview_text to design one from a description. Voice design also returns a preview clip
• Use your own voice right away: pass the returned voice name as voice to qwen-audio-3.1-tts-flash
• Playground: synthesize speech, create a voice and use it, all on one page

Notes

• Pricing: speech is billed by the input and output tokens actually used; creating a voice is a small flat fee per voice. Failed tasks are refunded in full. See pricing
• Sample requirements for cloning: WAV, MP3 or M4A from a public HTTP(S) URL, up to 60 seconds and 10 MB, 16 kHz or higher. 10–20 seconds of clear speech from one speaker works best
• Voices are account-bound: a custom voice works only with qwen-audio-3.1-tts-flash and only for the account that created it
• Voices are valid for 6 hours: a custom voice can be used for 6 hours after it is created; after that, create a new one
• Only clone voices you own or have explicit permission to use
• Try Qwen Audio 3.1 TTS Flash

Because the model vendor is retiring these models, four model IDs on EvoLink will be retired on October 9, 2026 (UTC+8). They will stop working from that date. Please switch to the replacement models before then.

Replacements

• qwen3-tts-vd → qwen-audio-3.1-tts-flash (speech synthesis)
• qwen-voice-design → voice-enrollment (custom voice creation)
• qwen-image-edit → qwen-image-3.0 (image editing)
• qwen-image-edit-plus → qwen-image-3.0-pro (image editing)

How to migrate

• Endpoints stay the same: audio is still POST /v1/audios/generations and images are still POST /v1/images/generations. Change model to the new model ID
• Voices must be created again: voices created with qwen-voice-design do not work with qwen-audio-3.1-tts-flash. Create them again with voice-enrollment. The parameter names are the same (voice_prompt, preview_text, preferred_name), with these limits: voice_prompt up to 500 characters, preview_text 15–200 characters, preferred_name 1–10 letters or digits. A new voice can be used for 6 hours after it is created. qwen-audio-3.1-tts-flash also includes 68 system voices, so you can synthesize speech without creating a voice
• Speech parameters: language_type is replaced by language (for example zh or en). The text limit per request rises from 600 to 5,000 characters. Unknown parameters return 400
• Image editing: reference images are still passed in image_urls
• Pricing: the new models are billed at their own rates. See Qwen Audio 3.1 TTS Flash pricing and Qwen Image 3.0 pricing
• View Qwen Audio 3.1 TTS Flash | View Qwen Image 3.0

GPT-6 Luna Decisions is now live on EvoLink: OpenAI's Decisions API on GPT-6 Luna. Send text or images with one or more questions and get typed answers (a probability, a fixed option, or a score) in one synchronous call.

GPT-6 Luna Decisions

• Model ID: gpt-6-luna-decisions; the OpenAI ID gpt-6-luna is also accepted on this endpoint
• Endpoint: POST /v1/decisions, on the same EvoLink API key
• Question types: predicate returns the probability that a condition is true, choice picks one of the values you supply, and score rates the input against ordered levels
• Input: text and inline base64 images in a user message, up to 128 image parts per request; hosted image URLs are not accepted
• Pricing: only input tokens are billed; output and cache tokens are free. Current rates are on the GPT-6 Luna Decisions model page
• Long context: requests above 272K input tokens bill the entire request at 2x the input rate
• Try GPT-6 Luna Decisions

Differences from Chat Completions

• Request fields: only model, input, questions, and the optional safety_identifier are accepted; stream, temperature, max_tokens, and other unknown fields return a 400 error
• No streaming, system prompts, or tools: for generated text, JSON in your own schema, or tool calls, keep using GPT-6 Luna on Chat Completions or Responses
• View pricing

Nano Banana 2.1 is now live on EvoLink: Google's update to Nano Banana 2 for image generation and editing at 1K, 2K or 4K, with web and image search grounding.

Nano Banana 2.1

• Model ID: gemini-nano-banana-2.1
• Endpoint: POST /v1/images/generations — text-to-image from a prompt, or editing and multi-image composition with up to 14 reference images in image_urls
• Resolution: quality accepts 1K, 2K and 4K. 0.5K is not available on this model
• Aspect ratio: size accepts auto or one of 14 ratios, from 1:1 to 21:9
• Search grounding: model_params.web_search and model_params.image_search are off by default; turn either on to let the model look things up before it generates
• Thinking level: model_params.thinking_level accepts auto, min, medium and high
• Pricing: per output image by resolution, with separate items for reference images, each search query the model actually executes, and text tokens. See pricing
• Async tasks: submit, then poll status or use callback_url. Result links are valid for 24 hours
• Try Nano Banana 2.1

deepseek-chat and deepseek-reasoner now redirect to DeepSeek V4.1 Flash on EvoLink, effective October 6, 2026 (UTC+8). Both legacy model IDs keep working with no changes to your API calls; for new integrations, use deepseek-v4.1-flash.

• Billing: both IDs are now billed at DeepSeek V4.1 Flash rates. Input and cache hits cost less than before; output costs more. Check the model pricing section for current rates
• Thinking: V4.1 Flash has thinking enabled by default. deepseek-chat requests that do not set a thinking option now run with thinking on, which adds latency and output tokens; set thinking.type to disabled to turn it off. Thinking tokens count toward max_tokens, so a small limit can be used up before any answer text is returned. With tool calls, echo the previous response's reasoning_content back in the follow-up request or the API may return 400; to skip this, turn thinking off for the whole session. See the API reference
• Limits: both IDs now get the V4.1 Flash context window of 1,000,000 tokens and a maximum output of 384,000 tokens
• Usage records: calls to both IDs appear under deepseek-v4.1-flash in your usage logs
• View DeepSeek V4.1 Flash | View pricing

Tencent Video Upscale is now live on EvoLink: send an existing video and get an enhanced copy at 720p, 1080p, 2K or 4K, billed per second.

Tencent Video Upscale

• Model ID: tencent-video-upscale
• Input: one MP4 or MOV video through a public URL in video_urls, up to 60 seconds and up to 4K. No prompt is needed
• Enhancement modes: mode picks one of seven — realistic (default), realistic_distant, animation, animation_distant, face_fidelity, detail_boost, soft. Every mode costs the same
• Target resolution: quality accepts 720p, 1080p (default), 2k and 4k. The target cannot be lower than the source video
• Frame rate: the output keeps the source frame rate and is never interpolated upward. model_params.max_fps (30, 60 or 120) sets a limit
• Pricing: per output second, by target resolution and by the frame rate of the delivered video; seconds round up to a whole second. See pricing
• Async tasks: submit, then poll status or use callback_url. Result links are valid for 24 hours
• Try Tencent Video Upscale

Kling 2.6 Motion Control is now live on EvoLink. Give it a character image and a motion video, and it returns a video in which the character performs the motion from the clip. It takes the same request as Kling 3.0 Motion Control at a lower per-second price.

• Model ID: kling-v2.6-motion-control (no aliases)
• Inputs: one character image (image_urls) and one motion video (video_urls); prompt is optional
• Orientation: model_params.character_orientation is required. image accepts motion videos up to 10 s, video up to 30 s
• Output: 720p or 1080p. Duration generally follows the motion video but may be shorter for complex or fast actions. The original audio is kept unless model_params.keep_sound is false
• Not supported: subject elements. Requests with a non-empty element_list are rejected; use Kling 3.0 Motion Control when you need it
• Pricing: billed per second of output video, rounded to the nearest whole second; 1080p is billed at a higher per-second rate than 720p. See pricing
• Switching versions: change model between kling-v2.6-motion-control and kling-v3-motion-control; remove element_list when switching to 2.6
• Async tasks: submit, then poll status or use callback_url
• Try Kling 2.6 Motion Control | View pricing | Compare the Kling family

Midjourney V8.2 joins EvoLink's unified image API with seven model IDs, Draft / Fast speed tiers, Standard / HD output and prompt-level --q detail levels that do not change the price.

Availability

Midjourney V8.2 is available through EvoLink's production API. Use the seven model IDs below with one EvoLink API key; the Midjourney V8.2 model page is the source of truth for live pricing and parameters. There are no aliases: the IDs below are the only names the API accepts.

Seven Model IDs

• mj-v8.2 — generate 4 images from a prompt (24 sketches in Draft), with image and style references
• mj-v8.2-variation — subtle or strong variations of one image from a completed mj-v8.2 task
• mj-v8.2-remix — regenerate a completed image with a new prompt while keeping its composition
• mj-v8.2-edit — place a completed image on a canvas and fill the rest from the prompt
• mj-v8.2-retexture — change materials and style of any public image URL while keeping its structure
• mj-v8.2-upload-paint — upload an image, mask a region and repaint it from the prompt
• mj-v8.2-remove-bg — cut out the subject of any public image URL, no prompt needed

Specifications

• Speed: draft, fast (default fast); Draft returns 24 low-resolution sketches at the same price
• Quality: standard or hd (default standard); Draft cannot be combined with HD
• Detail level: write --q 1 to --q 4 directly in the prompt; it is passed through unchanged and does not affect billing
• Prompt: up to 1024 characters of description; native parameters such as --ar, --s, --chaos and --sref are passed through
• Image input: image_urls must be public HTTP(S) URLs; retexture, upload paint and remove background take one image
• Derived tasks: variation, remix and canvas edit reference a completed mj-v8.2 task by task_id plus image_number (0–3 for a Fast task, 0–23 for a Draft task)
• Billing: per request, not per image; generation multiplies the base rate by quality (HD price ×1.5), edit routes are flat-priced, and every route is priced 10% below the official list price. See the pricing section for current rates
• API workflow: submit an asynchronous task to /v1/images/generations, then poll /v1/tasks/{id} for status and the image URLs

Model Highlights

• Draft to final in one key — explore composition with 24 sketches, then finalize the same prompt in HD
• Free detail levels — --q 2 to --q 4 add render time but never cost
• Edit chain on finished tasks — variation, remix and canvas edit start from a completed task ID, so nothing is re-uploaded
• Failed tasks are not charged — credits are held on submit and refunded on failure

Try Midjourney V8.2 | Pricing