Blend
Alloy scores every model on your own evals, then sends each request to whichever answers best under your latency and cost ceilings. Swap a model in the blend without touching a line of product code.
Alloy 2.0Routing engine now GA
Every model, melted into one endpoint.
Alloy blends 41 frontier and open-weight models behind a single API, and routes each request to the one that answers it best — inside the latency and price you set.
- Models blended
- 41
- Routing overhead
- 9 ms
- Tokens / month
- 3.1 T
- Uptime, 12 mo
- 99.995 %
The foundry
Built like metallurgy. Runs like water.
Temper
Guardrails, PII redaction and schema-validated outputs are applied at the edge in under 9 ms — identically for every model behind the endpoint, so safety never depends on the vendor of the day.
Cast
Distil your routed traffic into a private fine-tune. When it beats the blend on your evals, Alloy promotes it automatically, and rolls it back the moment it slips.
One call
Say what matters. Alloy picks the metal.
import { Alloy } from "@alloy/sdk";
const alloy = new Alloy({ apiKey: process.env.ALLOY_KEY });
const res = await alloy.respond({
input: "Summarise this contract and flag unusual clauses.",
route: {
optimize: "quality",
maxLatencyMs: 900,
maxUsdPer1k: 0.004,
},
fallback: "auto",
});
console.log(res.model); // "halcyon-3-large"
console.log(res.routeMs); // 8.6| Model | Score | p50 | $ / M tok |
|---|---|---|---|
| halcyon-3-large | 94.2 | 412 ms | 3.10 |
| kestrel-70b | 91.7 | 188 ms | 0.62 |
| lumen-vision-2 | 88.0 | 530 ms | 2.40 |
| ferro-mini | 79.4 | 74 ms | 0.09 |
Tighten the latency ceiling to 200 ms and the same call lands on kestrel-70b. No redeploy, no prompt rewrite, one invoice.
Pricing
Model usage at cost. You pay for the routing.
Ingot
$0/ month
For prototypes and weekend projects.
- 1M routed tokens a month
- 3 models per blend
- Community support
Foundry
$240/ month + usage
For products with real traffic.
- Unlimited blends and evals
- Edge guardrails and redaction
- 99.9% SLA, 4-hour support
Mill
Custom
For regulated and high-volume teams.
- Private fine-tunes from your traffic
- VPC peering, dedicated capacity
- 99.99% SLA, named engineer