Published Updated
In one or two sentences, before anything else.
WaveSpeedAI is an API platform, run by a Singapore company and a Hong Kong company named in its terms, that sells 1,000+ image, video, audio and 3D models plus an LLM service on one balance. It matches fal on Kling 3.0, charges above Google's list for Veo 3.1 Fast and accepts Alipay and WeChat Pay.
What it can do, one sentence each, and whether we actually tested it.
Image, video, audio and 3D models such as Seedance 2.0 and 2.5, Wan 3.0, Kling 3.0, Veo 3.1, Nano Banana Pro, Seedream 5.0 Pro, GPT Image 2 and MiniMax H3 behind one key.
Count from WaveSpeedAI's homepage; its blog claims early access to new ByteDance, Alibaba and Kuaishou models through direct partnerships.
Chat Completions and Responses at llm.wavespeed.ai/v1 and Anthropic Messages at llm.wavespeed.ai, for models from Anthropic, OpenAI, DeepSeek, Qwen and ByteDance; Claude Sonnet 5 is listed at $2 / $10 per million tokens.
WaveSpeedAI's LLM docs, which also cover prompt caching.
A hosted server at api.wavespeed.ai/mcp (OAuth or bearer key) and a local npx @wavespeed/mcp package, with tools to list and price models, run them, upload files and read balance, usage and billing.
WaveSpeedAI's MCP docs; a rerun of an unfinished job is billed again.
14 skills in WaveSpeedAI/agent-skills (a general skill plus per-model ones), and a Claude Code plugin marketplace, a Codex plugin and a Gemini CLI extension in the same GitHub organization.
Install with npx skills add WaveSpeedAI/agent-skills, per the repository.
Bronze (no top-up): 5 predictions a minute, 2 concurrent tasks. Silver (any top-up): 500 and 300. Gold (single top-up of $1,000 to $4,999): 3,000 and 3,000. Ultra ($5,000+): 5,000 and 10,000.
WaveSpeedAI's account-level docs and terms.
Credits do not expire, and failed requests, system errors and processing timeouts are refunded automatically.
WaveSpeedAI's terms; Trustpilot reviewers in April 2026 say failed or wrong-ratio generations were still charged.
Cards, PayPal, Google Pay, Apple Pay, WeChat Pay and Alipay, with auto-recharge, and wire transfer for enterprise.
WaveSpeedAI's pricing page; availability varies by region.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
FLUX.1 [schnell] and FLUX.1 [dev] images generated from the same scheduled prompts on WaveSpeedAI, fal and Replicate.
What happened
WaveSpeedAI was slower (FLUX.1 [schnell] 11.0 s with a p95 of 27.5 s, against 2.2 s on Replicate and 3.2 s on fal; FLUX.1 [dev] 14.9 s against Replicate's 4.2 s) but cheaper on FLUX.1 [dev]: $12 per 1,000 images, or $5 on its lossy version, against $25 on Replicate.
Case 2
Input
5-second, 720p videos from the same prompts on Vidu Q3 Pro and MiniMax H3 across providers.
What happened
WaveSpeedAI was the fastest on Vidu Q3 Pro (105.7 s against 117.1 s on Replicate and 117.8 s on fal, all at $9 a minute) but took 123.9 s on MiniMax H3 against fal's 12.5 s, at $6 a minute against $3.6. Video is sampled only a few times a day.
Case 3
Input
Claude Code started with ANTHROPIC_BASE_URL=https://llm.wavespeed.ai and a WaveSpeedAI key; Codex pointed at https://llm.wavespeed.ai/v1 with wire_api = "responses" in ~/.codex/config.toml.
What happened
One WaveSpeedAI key then covers the chat model, and its MCP server or skills can add media generation in the same agent. The guide warns that a successful Chat Completions request does not prove tool calls work with Codex.
What it does badly, with the evidence, and what we could not settle.
Weak
Slower than fal and Replicate on most shared image models.
Evidence: Artificial Analysis' three-day medians (October 21, 2026): FLUX.1 [schnell] 11.0 s against 2.2 s on Replicate and 3.2 s on fal; FLUX.2 [klein] 4B 6.2 s against 1.4 to 1.5 s. Its marketing promises sub-second images.
Veo 3.1 Fast above Google's own price.
Evidence: $0.15 a second with audio against Google's $0.10 at 720p, $1.20 against $0.80 for 8 seconds; Nano Banana Pro is $0.14 list against Google's $0.134 (WaveSpeedAI model pages and Google's Gemini API pricing, October 21, 2026).
Very low limits until you top up.
Evidence: The default Bronze level allows 5 predictions a minute and 2 concurrent tasks; any top-up lifts it to 500 and 300.
Terms that favour WaveSpeedAI.
Evidence: Suspension is possible immediately and without notice for 'commercial impracticality', total liability of all WaveSpeedAI entities is capped at US$100, and disputes go to SIAC arbitration in Singapore (terms, October 16, 2026).
Thin privacy terms.
Evidence: The privacy policy (September 25, 2026) does not address prompts or outputs, names no storage region or sub-processors and claims no SOC 2 or ISO certification; API logs are kept 'as needed'.
Billing complaints and little independent review.
Evidence: Trustpilot: four reviews, all one star, reporting a double auto top-up, charges for failed generations and a refused refund, with no company replies. A paying user's MCP bug report on GitHub went unanswered for about eleven months.
Not for teams that need the fastest image endpoints, strong contractual protection or clear data terms.
Evidence: fal and Replicate were faster on most shared image models, the liability cap is US$100, and the privacy policy is silent on prompts and outputs.
Uncertain
Minimum top-up amount.
Why unsure: Neither the pricing page nor the account-level docs state one.
Where data is stored and whether prompts or outputs train models.
Why unsure: The privacy policy names no region and does not mention training; the terms allow anonymized 'Resultant Data' to improve services.
Who funds the company.
Why unsure: A 'multi-million dollar' first round was announced in April 2025 with no amount or investors, and no later round was found.
Speed on proprietary video models.
Why unsure: Artificial Analysis covers open-weight and a few video models, not Veo, Kling or Seedance on WaveSpeedAI, and WaveSpeedAI's own speed claims come without a method.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | WaveSpeedAI PTE. LTD. (Singapore) and WaveSpeedAI LIMITED (Hong Kong), both named as providers in its terms of October 16, 2026 | About page: 3 Phillip Street, Singapore; founded 2025. Source: wavespeed.aichecked |
| Price against the maker's list | Veo 3.1 Fast with audio $0.15/s vs Google $0.10/s; Nano Banana Pro $0.14 vs $0.134 (4K $0.24, same as Google); no platform fee stated | Google's prices from its Gemini API pricing page, updated October 19, 2026. Source: wavespeed.aichecked |
| Minimum top up and account levels | Minimum not stated; any top-up lifts Bronze (5 predictions a minute, 2 concurrent) to Silver (500, 300) | Source: wavespeed.aichecked |
| Payment methods | Cards, PayPal, Google Pay, Apple Pay, WeChat Pay, Alipay (by region); wire transfer for enterprise | Source: wavespeed.aichecked |
| Free tier | $1 free credit for eligible new accounts, no card; some premium models excluded | Source: wavespeed.aichecked |
| Models | 1,000+ image, video, audio and 3D models; LLM service with Anthropic, OpenAI, DeepSeek, Qwen and ByteDance models | Source: wavespeed.aichecked |
| API formats | Media: REST tasks with polling or webhooks. LLM: OpenAI Chat Completions and Responses, Anthropic Messages; streaming and prompt caching documented | Source: wavespeed.aichecked |
| MCP | Official: hosted api.wavespeed.ai/mcp (OAuth or key) and local npx @wavespeed/mcp; models, prices, runs, uploads, balance and usage | Source: wavespeed.aichecked |
| Agent Skills | 14 skills in WaveSpeedAI/agent-skills (npx skills add WaveSpeedAI/agent-skills); plugins for Claude Code, Codex and Gemini CLI | Source: github.comchecked |
| Coding agents | LLM setup for Claude Code, Codex, OpenClaw, Claude Desktop, GitHub Copilot CLI and Gemini CLI; MCP setup also for Cursor and VS Code | Source: wavespeed.aichecked |
| Data retention | Generated media generally expires within 7 days; API usage logs kept 'as needed' | Source: wavespeed.aichecked |
| Status page | status.wavespeed.ai: API model health 99.966% on October 21, 2026; four incidents July to September 2026; homepage claims 99.99% | Source: status.wavespeed.aichecked |
| Refunds and liability | Credits never expire, non-refundable except at its discretion; failed tasks refunded automatically; liability capped at US$100 | Source: wavespeed.aichecked |
| Website visits, September 2026 (wavespeed.ai) | 1.77 million (−7.4% on August); 5th of 10 on AICPB's AI model routers list | Source: mp.weixin.qq.comchecked |
Answer first, then why, an example, the limits and what to do next.
WaveSpeedAI is a real business with public terms, a status page and official SDKs, but little independent review exists. Its terms name WaveSpeedAI PTE. LTD. in Singapore and WaveSpeedAI LIMITED in Hong Kong, and it is an official Hugging Face Inference Provider. Its four Trustpilot reviews are all one star, about billing and refunds.
Limits
Its terms cap total liability at US$100 and allow suspension without notice for 'commercial impracticality'.
WaveSpeedAI is not cheaper than Google on Google's own models. Veo 3.1 Fast with audio costs $0.15 a second against Google's $0.10 at 720p and Nano Banana Pro $0.14 per image against $0.134; Veo 3.1 at 1080p ($0.40 a second) and Nano Banana Pro at 4K ($0.24) match Google.
Limits
On open models it can be cheaper than rivals: Artificial Analysis lists FLUX.1 [dev] at $12 per 1,000 images on WaveSpeedAI against $25 on Replicate.
WaveSpeedAI was not the fastest on most shared image models in Artificial Analysis' measurements. Its three-day median for FLUX.1 [schnell] was 11.0 s against 2.2 s on Replicate and 3.2 s on fal (October 21, 2026); it was fastest on Vidu Q3 Pro video, at 105.7 s.
Why
Its own speed figures, such as up to 6x faster FLUX-dev on B200 GPUs, come from vendor posts without a defined baseline.
WaveSpeedAI gives eligible new accounts $1 of free credit without a card. Some premium models are excluded, and until you top up the account stays on the Bronze level of 5 predictions a minute and 2 concurrent tasks.
Limits
A Trustpilot reviewer (August 26, 2026) says the free trial led to a deposit demand.
WaveSpeedAI publishes an official MCP server, hosted and local. The hosted server at api.wavespeed.ai/mcp signs in with OAuth or an API key, the local one runs with npx @wavespeed/mcp, and both can list and price models, run them and read your balance and usage; setup is documented for Claude, ChatGPT, Claude Code, Codex, Cursor and VS Code.
The closest alternatives to WaveSpeedAI are fal, Replicate and Kie.ai. fal charges the same for Kling 3.0 and was faster on most shared image models; Replicate was faster on images but charges twice as much for Kling 3.0; Kie.ai is cheaper on Google's models but does not say how it gets access and has no status page.
This page is built from WaveSpeedAI's own site, docs, terms and status page, its GitHub repositories, Google's Gemini API price list, the model pages of fal, Replicate and Kie.ai, Artificial Analysis' provider benchmarks, and user reports on Trustpilot, GitHub and Hacker News, all checked on October 21, 2026. WaveSpeedAI's claims are labelled as its own.
According to AICPB's traffic rankings, wavespeed.ai had 1.77 million visits in September 2026, down 7.4% on August: 5th among AI model routers.
WaveSpeedAI sells one API key and one balance for what its homepage calls 1,000+ models: image, video, audio and 3D models such as Seedance 2.0 and 2.5, Wan 3.0, Kling 3.0, Veo 3.1, Nano Banana Pro, Seedream 5.0 Pro, GPT Image 2 and MiniMax H3. Its blog claims early access to new ByteDance, Alibaba and Kuaishou models through direct partnerships. A newer LLM service sells chat models on the same balance, for example Claude Sonnet 5 at $2 / $10 per million input / output tokens.
It began in inference acceleration and still markets speed. It also offers a desktop app, a CLI and ComfyUI nodes.
Eligible new accounts get $1 of free credit without a card. It takes cards, PayPal, Google Pay, Apple Pay, WeChat Pay and Alipay, the widest choice of the four media services. Credits never expire. The catch is the default tier: until you top up, a "Bronze" account is limited to 5 predictions a minute and 2 concurrent tasks. Any top-up lifts it to Silver (500 a minute, 300 concurrent).
Prices in US dollars, from each service's own model or pricing page and Google's Gemini API pricing (updated October 19, 2026). WaveSpeedAI's model pages warn that listed prices "may be outdated" and that the final task charge prevails.
| Model and output | Google's list | WaveSpeedAI | fal | Replicate | Kie.ai |
|---|---|---|---|---|---|
| Veo 3.1 Fast, 8 s at 720p with audio | 0.80 | 1.20 | 1.20 | 1.20 | 0.30 |
| Veo 3.1, 8 s at 1080p with audio | 3.20 | 3.20 | 3.20 | 3.20 | 1.275 |
| Nano Banana Pro, one 1K or 2K image | 0.134 | 0.14 | 0.15 | 0.15 | 0.09 |
| Nano Banana Pro, one 4K image | 0.24 | 0.24 | 0.30 | 0.30 | 0.12 |
| Kling 3.0 Standard, per second without / with audio | not confirmed | 0.084 / 0.126 | 0.084 / 0.126 | 0.168 / 0.252 | 0.07 / 0.10 (720p) |
| Seedance 2.0, per second at 720p | not confirmed | 0.24 | 0.3034 (with audio) | 0.18 | 0.205 |
On Google's models it is at or slightly above Google's list. Veo 3.1 Fast costs 50% more and Nano Banana Pro about 4% more at 1K or 2K, or 6% less when its 10% discount badge shows $0.126; Veo 3.1 at 1080p and Nano Banana Pro at 4K match Google. That makes it a little cheaper than fal and Replicate on Nano Banana Pro. Kling's and BytePlus's own prices could not be confirmed, so the last two rows compare the four services only: WaveSpeedAI ties fal on Kling 3.0 at half Replicate's price and sits in the middle on Seedance 2.0. Kie.ai is cheaper on most rows but does not say how it gets access to the models.
On open models it is often the cheapest: Artificial Analysis lists FLUX.1 [dev] at $12 per 1,000 images on WaveSpeedAI ($5 for its "Lossy Optimized" version) against $25 on Replicate. No platform fee is stated.
Artificial Analysis, an independent benchmarking firm, runs the same prompts on several providers every day. Its three-day median end-to-end times, read on October 21, 2026:
| Model | WaveSpeedAI | fal | Replicate |
|---|---|---|---|
| FLUX.1 [schnell] image | 11.0 s | 3.2 s | 2.2 s |
| FLUX.1 [dev] image | 14.9 s | not listed | 4.2 s |
| FLUX.2 [klein] 4B image | 6.2 s | 1.5 s | 1.4 s |
| Qwen-Image | 24.8 s | 40.1 s | 8.0 s |
| Vidu Q3 Pro, 5 s video at 720p | 105.7 s | 117.8 s | 117.1 s |
| MiniMax H3, 5 s video at 720p | 123.9 s | 12.5 s | not listed |
WaveSpeedAI was the slowest of the three on most shared image models, with a p95 of 27.5 s on FLUX.1 [schnell], which does not support its "sub-second" marketing on these endpoints. It beat fal on Qwen-Image and was the fastest on Vidu Q3 Pro video, at the same $9 a minute. On MiniMax H3 it matched MiniMax's own API (123.2 s) and took ten times as long as fal, at $6 a minute against fal's $3.6. Only open-weight and a few video models are covered.
WaveSpeedAI's own numbers are vendor claims: a joint post with DataCrunch (April 2025) claims up to 6x faster FLUX-dev inference against undefined "industry-standard baselines", and its blog posts comparing itself with fal and Replicate give no method. They read like neutral reviews but come from a competitor.
Trustpilot shows 2.6 out of 5 from four reviews, all one star, with no company replies. They report a "free trial" that turned into a deposit demand and a refused refund, reversed through PayPal (August 26, 2026); failed or wrong-aspect-ratio generations charged with support silent (April 23, 2026); auto top-up charging $55 twice despite a $15 setting (April 13, 2026); and content filters tightened without notice (December 2025). That is a small sample.
On GitHub, a paying user's September 2025 report that the MCP server did not work in Claude Code went unanswered until August 2026, when the CEO wrote that "eight months without a reply to a paying customer is not acceptable".
WaveSpeedAI publishes an official MCP server in two forms: a hosted server at https://api.wavespeed.ai/mcp, signed in by OAuth or a bearer key, and a local one run with npx -y @wavespeed/mcp@latest. Per its docs, the tools list and price models, run them, upload files, and read your balance, usage and billing. Setup is documented for Claude (web and Desktop), ChatGPT, Claude Code, Codex, Cursor and VS Code.
For Agent Skills, the WaveSpeedAI/agent-skills repository holds 14 skills, a general wavespeed skill plus per-model ones such as Nano Banana Pro and Veo 3.1 Fast, installed with npx skills add WaveSpeedAI/agent-skills; it names Claude Code, Codex, Cursor, OpenCode, Gemini CLI and others. Its GitHub also has plugins for Claude Code, Codex and Gemini CLI.
The LLM service speaks OpenAI Chat Completions and Responses at https://llm.wavespeed.ai/v1 and Anthropic Messages at https://llm.wavespeed.ai, with streaming and a prompt caching guide. Its coding-agent guide covers Claude Code, Codex, OpenClaw, Claude Desktop, GitHub Copilot CLI and Gemini CLI, not Cursor, and notes that a working Chat Completions call does not prove tool calls work in Codex. Official SDKs exist for Python and JavaScript, its GitHub organization has a Vercel AI SDK provider, and WaveSpeedAI is a Hugging Face Inference Provider.
fal charges the same for Kling 3.0 and a little more for Nano Banana Pro, and was much faster on most shared image models and on MiniMax H3. Replicate was faster on images and charges twice as much for Kling 3.0. Kie.ai is cheaper on Google's models but has no status page and does not explain its access. WaveSpeedAI's edge is payment choice, credits that never expire and one key for media and chat models.
Skip WaveSpeedAI if image speed matters most, if you need contractual protection beyond a US$100 liability cap, or if you send confidential inputs, since its privacy policy says nothing about them. Top up before real volume: the default tier runs two tasks at once.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Replicate is an API for running AI models, owned by Cloudflare since a $57.4 million acquisition closed on December 1, 2025. It sells 100+ official models such as Veo 3.1, Nano Banana Pro and Kling 3.0 per output, and 50,000+ community models by the second of GPU time.
fal is a San Francisco API platform, run by Features & Labels, Inc., that sells per-output access to more than 1,000 image, video, audio and 3D models on one prepaid balance. Its open models are cheap, but Google partner models such as Veo 3.1 Fast and Nano Banana Pro cost more on fal than on Google's own API.
Eden AI is a Lyon-based API gateway, run by Eden AI SAS, that puts about 500 language models and 250 'expert' models for OCR, speech, translation and image work behind one key and one prepaid balance. It passes provider prices through and, since January 2026, adds a 5.5% platform fee on self-serve credits.
CometAPI is a prepaid API service that sells text, image, video and audio models through one key at, by its own rule, 80% of the makers' list prices. Its terms and privacy pages name no company and use Hong Kong law, and it says it reaches the low prices through bulk credit buying and routing.