Platform
›LLM Tiering
LLM Tiering
The cost of the AI Layer is decided here and nowhere else.
Task → model
| Task | Model | Price (in/out per 1M) | Rationale |
|---|---|---|---|
| Policy Compiler | claude-opus-5 | $5 / $25 | Produces on-chain policy that executes money. Wrong output means lost funds — no saving here is worth it |
| Corporate Action Extractor | claude-sonnet-5 | $3 / $15 | High volume, structured extraction, output validated against a schema |
| Filing triage and materiality classification | claude-haiku-4-5 | $1 / $5 | Thousands of filings a day, mostly noise. Filter cheap first; escalate only what passes |
| Risk Explainer | claude-haiku-4-5 | $1 / $5 | Paraphrasing signals into human sentences. Light task, latency-sensitive |
Prompt caching is mandatory
Roughly 0.1x on a cache read
The system context — token list, schemas, thresholds — is identical across requests. At every tier, caching it is not an optimization to revisit later; it is the difference between a viable and an unviable cost structure.
On This Page