Adaptive routing
Send each request to the model that answers it best, scored on live quality and cost rather than a fixed preference order.
38%median cost reduction
The platform
One control plane for every model you run
Routing, evaluation, caching and observability in a single layer, so switching providers is a config change rather than a migration.
Platform
Everything between the request and the answer
Six systems that most teams end up building badly. They are the product, not add-ons, and every plan includes all of them.
Semantic caching
Cache on meaning, not on string equality, so paraphrased questions hit the same entry and never reach a model.
4.1xeffective hit rate
Continuous evaluation
Every deployment replays your graded set before it takes traffic, and rolls back on its own if quality drops.
2,400graded cases per run
Structured pipelines
Compose retrieval, tools and generation as a typed graph. Each node is independently versioned, cached and traced.
11msorchestration overhead
Managed retrieval
Chunking, embedding and reranking behind one endpoint, with incremental reindexing as your corpus changes.
92%recall @ 10
Policy enforcement
Redaction, jailbreak detection and per-tenant rate limits applied at the edge, before a prompt reaches a provider.
0prompts stored by default
Routes
Every model behind one endpoint
Add a provider or your own checkpoint and it becomes a candidate route immediately. Latencies are illustrative first-token figures.
| Model | Vendor | Context | First token | Strengths |
|---|---|---|---|---|
| Meridian 3 Opus | Meridian | 400k | 480ms | ReasoningLong context |
| Meridian 3 Flash | Meridian | 200k | 120ms | ThroughputCost |
| Kestrel L | Northwind | 128k | 260ms | CodeTools |
| Kestrel S | Northwind | 64k | 90ms | Classification |
| Atlas Vision | Atlas | 96k | 540ms | VisionDocuments |
| Ember 7B | Self-hosted | 32k | 40ms | PrivateEdge |
How it fits
A day to adopt, not a quarter
Halcyon is a base URL and a key. Point your existing client at it and nothing else has to change on the first day.
Drop-in compatible
Speaks the API your SDK already speaks. Change the base URL and your calls keep working.
Versioned everything
Prompts, pipelines and routing policies are versioned artifacts you can diff, review and roll back.
Bring your own models
A self-hosted checkpoint is just another route, scored and cached like any hosted one.
Private by default
Nothing is retained unless you turn retention on. Redaction runs before a prompt leaves your VPC.
Fails to a good answer
A provider outage reroutes mid-flight. Your users see a slower response, not an error page.
One bill
Every provider, your own capacity and the platform on a single invoice, broken down per feature.
Contact
Tell us what you are building
An engineer reads every message. If you are moving an existing workload across, mention the provider and we will send the migration notes for it.
engineering@halcyon.dev
Docs
docs.halcyon.dev
Status
status.halcyon.dev
Office
Lisbon, remote-first