This page is built from AI/ML API's own site, docs, terms, privacy policy and status page, the Estonian company register, the official price pages of OpenAI, Anthropic and Google, Trustpilot reviews, GitHub issues and Hacker News, all checked on October 21, 2026. AI/ML API's claims are labelled as its own.
What AI/ML API sells
AI/ML API (also written AIMLAPI) sells one prepaid key for other companies' models. Its pricing page claims "1000+ models" across text, image, video, voice, music, embeddings and 3D, including GPT-6 Sol, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7, DeepSeek V4, Sora 2, Veo 3.1, Kling and Flux. Third-party directories have quoted 100+, 200+, 300+ and 400+ at different times. The balance is kept in credits at 1 USD = 2,000,000 credits, and Pay As You Go starts with a 20 USD top-up. The docs FAQ says "The Free Tier is currently paused", so there is no way to try the API before paying.
The API is OpenAI-compatible at https://api.aimlapi.com/v1. Claude Code works by pointing ANTHROPIC_BASE_URL at https://api.aimlapi.com, and a Batch API accepts Anthropic-style batches for Claude models. The docs describe no provider selection, fallback or load-balancing settings: AI/ML API picks the upstream.
Prices against the official APIs
For the current OpenAI, Anthropic and Google models checked, AI/ML API's prices are 1.3 times the makers' list prices. Some older Claude models are 1.375 times.
| Model, per million input / output tokens | AI/ML API | Maker's list price | Source |
|---|
| GPT-6 Sol | 2.60 / 13 USD | 2 / 10 USD | OpenAI |
| GPT-5.4 | 3.25 / 19.50 USD | 2.50 / 15 USD | OpenAI |
| Claude Opus 5.5 | 5.20 / 26 USD | 4 / 20 USD | Anthropic |
| Claude Sonnet 5.5 | 2.60 / 13 USD | 2 / 10 USD | Anthropic |
| Claude Sonnet 4.6 | 4.1262 / 20.631 USD | 3 / 15 USD | Anthropic |
| Gemini 4 Argon | 2.60 / 13 USD | 2 / 10 USD until November 30, 2026; 4 / 20 USD from December 1 | Google |
AI/ML API prices come from its pricing page, which uses "$" without naming a currency. There is no separate platform or deposit fee; the margin is inside each price. Older directory copy repeats a vendor claim of saving "up to 80%" against OpenAI, which the current price list does not support. Video is priced per second (Sora 2 at 0.13 USD a second, Veo 3.1 at up to 0.78 USD a second); OpenAI's current pricing page no longer lists Sora, so that markup could not be compared.
What public tests show
No third-party latency, throughput or output-quality test of AI/ML API was found in public sources. What exists:
- AI/ML API's own status page (UptimeRobot, live on October 21, 2026) showed the API at 100.000% over 30 days and 99.980% over 90 days, the MCP server at 99.959% and 99.962%, and one 20-second outage on October 14, 2026. These are endpoint pings, not per-model success rates.
- A Trustpilot reviewer (November 1, 2025) says technical tests showed a model sold as GPT-5 answering as GPT-4, and that a refund was refused. AI/ML API replied that model misidentification is a known issue across providers and that it does not misrepresent models. The method was not published, and asking a model its name is an unreliable test.
- AI Indigo, a directory site, rates it 2.7 out of 5 without an author or method.
What users report
Trustpilot shows 3.4 out of 5 from 23 reviews (October 21, 2026), split between 52% five-star and 44% one-star. Praise covers the model range, same-day model additions, the playground and quick support. Complaints cover refused refunds (including an accidental 20 USD charge), a dashboard that showed over 135 USD for nine requests (the company called it a display bug), no way to test before paying, timeouts and "model overloaded" errors. Trustpilot notes the company replied to 75% of negative reviews, usually after more than a month.
Most Hacker News mentions are Show HN posts and comments by AI/ML API staff accounts. On GitHub, a staff member opened issues on open-source projects such as AutoGPT and Activepieces in August and September 2026, offering a "50/50 revenue share" for adding AI/ML API as a provider and claiming more than 400,000 users. Some provider integrations in open-source tools therefore come from paid outreach.
MCP, Agent Skills and coding agents
AI/ML API runs an official remote MCP server at https://mcp.aimlapi.com/mcp. Clients such as Claude Desktop, claude.ai, Cursor and Claude Code connect by browser sign-in (OAuth) or with an API key, and can then search and compare models, run text, image, video, audio and embedding models, manage long jobs and check the balance. The same server exposes its Toolbox: four tools (tools_search, tools_inspect, tools_run, tools_run_status) that reach web search, page-to-Markdown, screenshots, news and social data. Only tool runs are billed. Its Toolbox docs also show the MCP server passed as an mcp tool in a /v1/responses request for OpenAI models, so the model calls the tools inside one request.
For coding agents, the aimlapi CLI (npx aimlapi claude) signs you in, creates a key and starts Claude Code, Codex, OpenCode, Cline, OpenClaude, OpenClaw, Hermes Agent, zero or ZCode already pointed at AI/ML API. The docs also have guides for Roo Code, Kilo Code, Aider and continue.dev; Cursor appears only in the MCP guide. The Claude Code guide warns that only the Claude models AI/ML API carries work through the Anthropic format.
No SKILL.md skill repository was found. The docs publish an "Agent setup skill" page, which is a set of instructions for a coding agent to install the CLI. Official SDKs are aimlapi-sdk-python (built on the OpenAI SDK) and aimlapi-sdk-node, plus the langchain-aimlapi package; there is a LiteLLM guide and an llms.txt index of the docs.
Privacy, terms and who runs it
The terms name Boiler Labs FZ-LLC, in a Ras Al Khaimah (UAE) free-zone business centre, as the contracting party. The site footer also lists AIMLAPI OÜ, an Estonian private company entered in the register on February 29, 2024 with a share capital of 1 EUR, and the terms put disputes under Estonian law and the Harju District Court. AIMLAPI OÜ reported revenue of 260,860 EUR for 2025 with a loss of 22,765 EUR, and 109,460 EUR for 2024. No public document explains how the two companies relate or which one receives payments.
- Stripe is merchant of record, and the terms say the company "does not offer refunds for any products or services".
- The terms allow fair-use rate limits and immediate suspension without notice, and cap liability at the amount paid.
- The home page advertises a "99.9% Uptime SLA"; the terms disclaim any guarantee of uptime or reliability.
- The privacy policy has a GDPR section naming Boiler Labs as controller, but says nothing on whether prompts and outputs are logged, for how long, used for training or sent to which upstream providers. No SOC 2 or ISO certification is named.
How it compares with other relays
| Service | Price against makers' list prices | Fee on top |
|---|
| AI/ML API | About 1.3 times | None stated |
| OpenRouter | Same | 5.5% on credit purchases |
| CometAPI | 0.8 times, by its own rule | None stated |
| ZenMux | Same | 5% on top-ups, waived for now |
| NanoGPT | Same | None on deposits |
OpenRouter also lets callers choose and order providers. AI/ML API's case rests on breadth (video, music, 3D and voice under one key), its MCP server and Toolbox, and the one-command CLI.
Who should skip it
Anyone sensitive to price, anyone who may need a refund, and teams that need written data-retention terms, a contractual SLA or knowledge of which upstream serves each request. For trying many media models from one key with an MCP server attached, the main costs are the 30% premium and a non-refundable 20 USD top-up.