Apply now
If your work requires LLM inference, have observed venue inefficiencies that want to improve, or want a microsecond-overhead routing with founder-level support, reach out to us. You'll work directly with the founder:
You only need to point your OpenAI client at Kottos AI and every request is automatically recorded: which venue, what cost, how long it took, and whether it failed.
# Currently BYOK: your keys are passed per request and never stored. client = OpenAI(base_url="https://api.kottos.ai/v1", api_key=KOTTOS_KEY)
Four things to consider to reduce your LLM inference cost.
For open-weight models, several providers can offer inference at different rates. For instance, Llama-3.3-70B input price goes from about $0.10 to $1.04, see live prices below. Depending on your quality floor, you can optimize which is the cheapest provider that meets your needs. But this is not the only way to reduce your bill. On example is the cache lifetime config. The five-minute config was 36% more expensive than the one-hour config on our own Claude traffic.
Claude Fable 5.1 quoted prices are $10 in, $50 out per million tokens, with $0.25 for cache reads. We computed our effective cost on our Claude Code traffic and came out to be $0.54 per million tokens at Anthropic's list prices, because of the cache reads. You can get this number automatically from the tape.
Knowing when to switch providers is important. We simulated a price-weighted routing model with two venues sharing the same quote price on 6,361 requests from our own traffic. Under this simulation, the token costs rose from $1,360 to $1,646, which is a 21% increase in cost.
We benchmarked the OpenAI service tiers for GPT 5.6 Terra on 09/17/2026 and found out that the standard tier was 3x slower at the median than the cheapest, and 2x the price. Note that this is a measurement which can change as demand changes. This is why Kottos AI intelligent routing exists.
You only point your client to Kottos AI and we handle the rest.
Kottos AI uses quotes and executed requests to learn the optimal routing for the specific task. How routing uses the tape →
We handle translation from an OpenAI-compatible API to: OpenAI, Anthropic, Google, Bedrock, Together, Fireworks, Groq, and more. You can change providers without touching your code.
We provide production observability and monitoring. This can be used for cost attribution by user, team, and project.
You work with the engineering team directly. We welcome custom integrations.
Open-weight models can have many providers serving them at different prices. There is a routing opportunity but not without a catch. It is important to observe that while providers serve the same model, they do not necessarily offer the same product: different quantization, context limit, latency, or uptime. As of 07/28/2026, Llama-3.3-70B input prices go from $0.10 to $1.04, where the cheapest is usually the one with lower quantization or context. So routing is picking the cheapest that meets your task requirements. Live prices:
| Provider | Input $/M | Output $/M | Quant | Context | Uptime |
|---|---|---|---|---|---|
| Fetching live prices from OpenRouter... | |||||
Fetched live from OpenRouter's public API.
llmbridge is the open-source fast gateway core engine. It runs translation, offers a single-provider proxy, and is benchmarked to produce a microsecond overhead. This does not contain any routing or price-performance intelligence.There are different costs we can base the routing on. For latency routing, we look at the TTFT of each request and route to the current fastest venue. For cost routing, we look at the expected cost we predict and decide whether a venue is the optimal option. To switch venues, we need to cross a threshold of minimum gain. See more about what is available today in the routing page.
No. We never train on customer requests or responses text. We collect anonymized latency metadata to improve routing. We only use prompt and completion content to serve the request.
If your work requires LLM inference, have observed venue inefficiencies that want to improve, or want a microsecond-overhead routing with founder-level support, reach out to us. You'll work directly with the founder: