Apply now
If you're running production LLM workloads and want microsecond-overhead routing with founder-level support, send a brief note about what you're building. You'll work directly with the person who wrote the hot path:
Point your OpenAI client at Kottos AI and every request is recorded: which venue served it, what it cost, how long the first token took, whether it failed. The price book runs live across providers today. The latency and error record is kept on every hosted request, and routing consults it. Bring your own provider keys. No token markup, ever.
# Any OpenAI SDK. Change the base URL and the key. The code stays. client = OpenAI(base_url="https://api.kottos.ai/v1", api_key=KOTTOS_KEY) # Your provider keys stay yours: passed per request, never stored in the request path.
Four things, and each one is a number you can check.
Llama-3.3-70B input runs from about $0.10 to $1.04 per million tokens across the providers that serve it, live from the price book below. Routing captures the gap among the venues that meet your quality floor, not the cheapest overall. Our 15–40% savings figure is an estimate from that dispersion. Measured savings get published as design partners go live.
Claude Fable 5.1 lists at $10 in, $50 out per million tokens direct from Anthropic, with cache reads at $0.25. Our own Claude Code traffic through the gateway came to a blended $0.70 per million tokens at Anthropic's list prices, in and out, across its Fable 5.1 requests (September 2026), because most of the input was cache reads. Nothing on a price page tells you that. The tape does, per request.
We replayed 6,361 Claude Opus 5 requests of our own traffic through the rule the public routers document, flow split by posted price, over the two venues that sell it at the same list price. Every switch lands on a venue that has not seen the prompt, so cache reads at $0.50 per million become writes at $10. The replay priced $1,646 against $1,360 at list, 21% more with no price gain, because the quotes were identical. On five-minute cache entries the same replay reads 39% more. The tool is ours, the prices are Anthropic's list, the traffic is ours, and it measures a spread, not a saving. We were not first to publish the finding: Factory AI, who build enterprise coding agents, published the same argument in August 2026 from their own production traffic, reporting that routing without preserving the warm prefix would have cost 2.12× their baseline.
A customer benchmarked the three service tiers of one OpenAI model. We measured the wait before the first response byte, which is where a service tier is bought or sold. Both times the flex tier, the cheapest of the three, had the fastest median.
Tokens pass through at the provider's own rate. Our fee for the routing service is quoted per pilot and is never a percentage of what you spend on tokens, so it does not rise because you picked a costlier model. A percentage charges you more for identical routing work.
“Hosted” means we run the gateway for you, but that is not the only option. The same engine and the same routing intelligence are available with the data plane inside your own network.
Point your existing OpenAI client at our endpoint and you are done. We run the gateway and we route: every request takes the best venue our measurements identify among those you can reach. Observability, team management and infrastructure included. Nothing to deploy, nothing to operate.
The same Kottos gateway, deployed as a Docker container inside your network, so prompts and provider keys never leave it and the gateway sits next to your app for full sub-millisecond latency. It connects out to the Kottos control plane for real-time routing intelligence and usage metering, and you choose: our book recommends the venue, your gateway executes (or overrides) it. The tape itself stays managed by us.
Separately, the open-source core. llmbridge, the translator and proxy, is self-hostable under Apache 2.0 with no account and no connection to us at all. It gives you the speed, but none of the multi-provider routing, price book or observability: that intelligence is the commercial layer.
Multi-region deployment, automatic failover, and an uptime SLA. You point your client at our endpoint, and we operate the gateway and handle the rest.
The price side of the tape runs live today: the same open-weights model tracked across providers in real time, priced up to ~10× apart (≈10× for Llama-3.3-70B right now). The latency/error side is measured on every hosted request, a rolling p95 per venue that routing consults. The cross-provider comparison is in build with design partners, routed by us when we host, recommended to your gateway when you do. We estimate 15–40% lower effective cost, measured per customer as design partners go live.
One OpenAI-compatible API across OpenAI, Anthropic, Google, Bedrock, Together, Fireworks, Groq, and more. Switch providers without touching client code.
Per-request OpenTelemetry tracing. Token-cost attribution by user, team, and project. Budgets, forecasting, and audit logs. Prompt content is never logged.
You work with the engineering team directly: same-day response, custom integrations welcome.
OpenAI-compatible surface. Change your base URL and API key, and existing LiteLLM / OpenRouter / Portkey code works. Most teams migrate in under a day.
The routing opportunity, made concrete, and the catch. The same open-weights model is sold by 15+ providers, but not as the same product: they differ in quantization (fp8 vs fp16, which is output quality), context limit, latency, and uptime. As of 07/28/2026, Llama-3.3-70B input runs about $0.10/1M on the cheapest provider versus ~$1.04 on the priciest (up to ~10×) and the cheapest is usually the quantized / smaller-context one. So routing isn't "pick cheapest". It's pick the cheapest that meets your quality, latency, and error budget. Live price and specs, side by side:
| Provider | Input $/M | Output $/M | Quant | Context | Uptime |
|---|---|---|---|---|---|
| Fetching live prices from OpenRouter... | |||||
Fetched live from OpenRouter's public API when this page loads. Figures are OpenRouter's per-provider rates (mostly provider pass-through), not a Kottos quote.
There are three ways to run Kottos AI, and they differ on who operates the data plane, not on how fast it is. The engine is the same in all three.
llmbridge, the fast gateway core: format translation and a single-provider proxy, microsecond overhead, Apache 2.0, no account. It is deliberately "dumb": there is no routing, no price-performance intelligence, no tape. Free forever.Put simply: the open-source core gives you the speed, and the hosted or self-hosted plans add the intelligence. The difference between them is only where the data plane runs. llmbridge is free. The hosted and self-hosted plans charge for the routing service, quoted per pilot, never a percentage of token spend and never a token markup.
The price side of the tape runs live today: the same model tracked across every venue that sells it. The latency and error side is measured on every hosted request: the tape records what each venue delivered, and routing consults a rolling p95 per venue. What is still in build, with design partners, is the comparison of the same model across providers. On each request we select from the venues available to you, choosing the provider that's actually fastest and cheapest right now for your model and constraints, not the lowest list price. The same model can be priced up to ~10× apart across providers (≈10× for Llama-3.3-70B today), so the headroom is real. Our 15–40% figure is an estimate from that dispersion, and we'll publish measured savings as design partners come online.
No. We never train on customer requests or responses. We collect anonymized latency metadata to improve routing. Prompt and completion content is never logged or used for anything beyond serving the request.
Yes. This is the self-hosted model: deploy the Kottos gateway as a Docker container in your own VPC, so prompts and provider keys never leave your network and it sits next to your app for full sub-millisecond latency. It connects to the Kottos control plane for real-time routing intelligence and usage metering, the tape itself stays managed by us. The open-source llmbridge core (translation + proxy) is separately self-hostable under Apache 2.0, no account needed.
If you're running production LLM workloads and want microsecond-overhead routing with founder-level support, send a brief note about what you're building. You'll work directly with the person who wrote the hot path: