The fastest way to ship with the latest open-weight models

AI Token Router is an OpenAI-compatible API gateway for open-weight AI models. Point your existing OpenAI SDK at one endpoint and you get Kimi K3, GLM-5.2, DeepSeek V4 and open-source video models like Wan 2.2 and LTX-2.5, at 30-50% below official pricing. One key, one bill, and the official rate printed next to ours on every row.

We buy inference capacity in bulk, run it at high utilisation, and pass most of that discount through instead of keeping it as margin. There is no platform fee and no surcharge on top of the per-token rate you see. It is also why we only carry open-weight models: their licenses permit third-party hosting, so we control the cost base rather than reselling someone else's API at a markup.

No. The API is OpenAI-compatible, so if you already use the OpenAI SDK -- or any framework built on it, which is most of them -- switching is a one-line change to your base URL. Request shapes, streaming frames, the usage block and the error envelope all match what your code already handles.

Open-weight text models including Kimi K3 and K2.6, GLM-5.2 and GLM-5.2 Air, DeepSeek V4 Pro and Flash, Qwen3, MiniMax M2, Llama 4, Mistral Large 3 and GPT-OSS. On video: Wan 2.2 text-to-video and image-to-video, LTX-2.5, HunyuanVideo 1.5, CogVideoX and Mochi 1. On images: FLUX.2 schnell, Qwen-Image, Stable Diffusion 3.5 and HiDream-I1. New open-weight releases go live within 24 hours at a published price.

When you resend the same prefix -- a long system prompt, a fixed tool schema, a document you keep querying -- it is served from the model's KV cache and billed at the cached-input rate, typically 80-90% below the standard input rate. It happens automatically with no header to set, and a cache entry lives about five minutes after its last use. For agent loops this is usually the largest line on the bill, which is why the cached rate has its own column on the pricing table instead of a footnote.

No. The per-model rate is the whole price. No platform fee, no minimum spend, no card-processing surcharge at checkout, and no charge for failed requests. Worth checking when you compare: OpenRouter adds roughly 5.5% on credit purchases, which does not appear on their per-model pricing table.

Your prepaid balance is a hard ceiling, enforced by the gateway on every request — a runaway agent stops at what you have already paid, never at an invoice. Unlike a configured cap, it cannot be misconfigured. Automatic recharge is off by default, and when enabled it is bounded too: a fixed amount, capped at three recharges in any 24 hours, switched off after three consecutive declines. Per-key limits are on the roadmap.

Eligible requests fall back to another available model rather than failing outright, and the fallback is recorded in your request log so you can see when it happened and what it cost. You can also pin a request to a single model when substitution would be wrong for that workload.

We do not train on your prompts or completions, ever. By default request bodies are retained for 30 days so you can debug from the request log, and you can turn that off in settings for zero data retention, where prompts and completions are discarded as soon as the response is returned. Usage metadata -- model, token counts, latency, cost -- is always kept, because that is what the dashboard and your invoice are built from.

Yes. If you already have a contract with a provider, connect that key and route through us anyway. You keep your negotiated rate and still get routing, request logs, spending caps and per-key cost attribution. There is no charge for BYOK traffic.

OpenRouter is a router across 500+ models including closed frontier models like GPT, Claude and Gemini, with multi-provider failover. If you need those, OpenRouter is the better fit and we say so on our comparison page. We are narrower on purpose: open-weight models only, roughly 22 of them chosen deliberately, with no platform fee, published cached-input rates, and open video generation treated as a first-class category rather than something we do not carry.

Because reselling closed-weight model access frequently violates the origin provider's terms of service, and that is a risk we will not take on or expose customers to. Open licenses permit third-party hosting, which is what lets us run the inference ourselves, control the price, and publish it. It is also why we do not serve Sora 2, Veo 3.1, Runway Gen-4, Kling, or Wan 2.5/2.6 -- the last of which is API-only, with weights never released, despite the open Wan 2.1/2.2 lineage.

Yes. The Playground gives you 3 free runs with no account and no credit card, using representative sample responses so you can see the request and response shape before signing up. Live inference against the models starts the moment you have a key. Every model page also carries a copy-paste-runnable snippet with the model id pre-filled.

No. Sign up with GitHub or Google -- no password to set, nothing to leak -- and you get $5 in credits plus 2M free tokens on any open-weight model. You only add a payment method when the free credit runs out, and there is no subscription -- it is pay as you go with no minimum spend, cancel anytime.

Yes. Accounts spending $1,000+ per month qualify for custom rates, with deeper discounts, priority support and dedicated SLAs at higher tiers. We reply within 24 hours.