Eden AI pricing: 5.5% fee, free tier, vs OpenRouter

This page is built from Eden AI's website, pricing and model pages, docs and llms.txt, terms, status page and GitHub, Anthropic's, OpenAI's and Google's price pages, a pricing-change tracker, and user reports on Hacker News, Trustpilot, Product Hunt and G2, all checked on October 21, 2026. Eden AI's claims are labelled as its own.

What Eden AI sells

Eden AI started in 2021 as one API for many machine-learning services and now sells two things behind one key and one prepaid balance. The first is an LLM gateway: an OpenAI-compatible endpoint at api.edenai.run/v3 for models from OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, Qwen and others, with ids such as anthropic/claude-sonnet-latest. The second is what it calls expert models: OCR, document and resume parsing, speech-to-text, text-to-speech, translation, image analysis, web search and video generation from many providers, through a separate /v3/universal-ai endpoint.

Its pages give different counts: about 1,000 models on the home page, 500+ in the docs, and 750+ (500+ language models plus 250+ expert models) on the models page.

A request can list fallbacks (up to 7 LLMs or 3 expert models) and fail over between providers of the same model, with at most 8 attempts, billed only for the one that answers. An EU endpoint, api.eu.edenai.run, sends requests only to EU-eligible providers and refuses others with HTTP 451, without billing them.

Prices against the official APIs

Eden AI says it passes provider prices through and adds a 5.5% platform fee on the self-serve plan. For the models checked, its catalog matched the makers:

Eden AI's prices come from its models page. With the 5.5% fee, 100 USD of model usage costs 105.50 USD; on AWS Marketplace a 1,000 USD credit pack costs 1,055 USD.

The fee is new and described two ways. A PricingSaaS capture shows it was added between January 19 and 27, 2026; before that Eden AI earned its margin from suppliers. Eden AI's pricing and plan docs say the fee is applied at checkout when you buy credits, but its own OpenRouter comparison post says it applies only to actual API usage. The BYOK doc says requests on your own keys are billed by the provider and mentions no Eden fee, but BYOK needs "a plan that includes BYOK". There are no free credits; sandbox tokens let you test an integration without real provider calls, at no cost. The terms say payments are non-refundable unless otherwise stated.

What public tests show

No third-party benchmark of Eden AI's latency, routing or uptime has been published. What exists:

• Hacker News user muzzy19, who submitted the April 26, 2026 thread, said after about a week of use that it was "working fine", and valued the European models and pay-as-you-go billing. No numbers.
• An edX product manager wrote on Product Hunt, around 2023, that its text APIs worked well against forum spam. That predates the LLM gateway.
• Eden AI's own blog compared six speech-to-text providers through its API on one audio file per use case, and warns that real projects need a representative sample.
• A PromptZone article quotes 200 to 500 ms for basic queries from unnamed "community reports", with no method.
• Eden AI's home page claims 99.99% uptime; its status page showed all systems operational and no incidents in the past seven days.

What users report

Eden AI reached the Hacker News front page on April 26, 2026 as "Eden AI - European Alternative to OpenRouter" (133 points, 69 comments). Commenters questioned the "European" label when most models on offer are American or Chinese, noted that OpenRouter also charges 5.5%, flagged a missing imprint and the non-refundable payment clause, and said the router itself offered no zero-retention agreement. Others said they would switch at similar prices and that one contract and one invoice was the real value. The terms page now names Eden AI SAS and its address.

Ratings are thin. Trustpilot shows 2.3 out of 5 from 7 reviews, but its one-star reviews either name a different product, an AI companion app, or do not say what they refer to. Product Hunt shows 5.0 from 21 reviews, mostly from 2023, and G2 has one review.

MCP, Agent Skills and coding agents

Eden AI runs an official hosted MCP server at mcp.edenai.run/mcp (Streamable HTTP, Bearer key). It publishes one tool per expert feature, such as ocr, web_search, image_generation and translation, plus upload_file, check_job and list_models. Each tool call is billed as a normal Eden AI call under the same rate limits and data rules. The docs show setup for Claude Code, Codex CLI, Cline, Continue, OpenCode, Hermes Agent and OpenClaw, and how to hand the tools to any function-calling model without an MCP client. Passing MCP servers to models through the LLM endpoints is not documented.

Eden AI also publishes an official Agent Skill, edenai/edenai-skill on GitHub, created in April 2026. It teaches Claude Code, and through the skills CLI agents such as Cursor, Codex and OpenCode, to use Eden AI for LLM chat, OCR, speech, web search, its MCP server and account tasks.

Coding agent guides cover Claude Code, Codex CLI, Cline, Continue, OpenCode, Kilo Code, Pi, Hermes Agent and OpenClaw; Cursor, Roo Code and Zed have none. For Claude Code you set ANTHROPIC_BASE_URL to https://api.edenai.run/v3 and ANTHROPIC_API_KEY to your Eden AI key, and can switch to models such as openai/gpt-latest.

The API covers OpenAI chat completions with streaming, tool calling, structured output, web search and prompt caching; the OpenAI Responses API; an Anthropic Messages pass-through with token counting; and the universal endpoint for expert models. There are guides for the OpenAI SDKs, LangChain, Pydantic AI and Haystack, a LiteLLM provider, an n8n node and an llms.txt.

Privacy, terms and who runs it

Eden AI SAS is registered in Lyon (RCS 829 046 473) with an address in Villeurbanne, France, and is led by co-founder and CEO Taha Zemmouri. It raised a 3 million EUR seed round led by Galion.exe in late 2024. Eden AI says prompts and outputs are not stored by default, log retention is opt-in per project, async results are kept seven days, and it does not train on customer data. A DPA is available, and the site claims SOC 2 and ISO 27001. Data may leave the EEA depending on the providers you choose, unless you use the EU endpoint. French law governs the terms, and liability is capped at fees paid in the previous 12 months. The self-serve rate limit is 10 requests per second across keys (another page says 7), raised on request.

How it compares with OpenRouter, Vercel and Cloudflare

• OpenRouter charges the same 5.5% on credit purchases on its Standard plan and lists more open-weight hosts and free models, but no OCR or translation APIs.
• Vercel AI Gateway bills at list price with no fee, and covers image and video models but not OCR or translation.
• Cloudflare AI Gateway adds 5% to credit purchases and caps those requests at 200 a minute per gateway.
• Requesty adds 5% to each top-up.

Who should skip it

Skip it if you want the cheapest tokens, a refund if things go wrong, or high request rates from day one. For LLM-only use it costs the same as OpenRouter and more than no-fee relays. It makes most sense for EU teams that want LLMs plus OCR, speech and translation on one invoice, with a DPA and an EU-only endpoint.