Milliseconds, not seconds
Myles-G decides in tens of milliseconds on one GPU; Myles-22M in about 27 ms in a browser.
Solutions · AI product teams
The decisions inside your agents, in milliseconds.
Every AI product is full of small decisions: is this input safe, which tool should the agent call, is this answer grounded, should a person check this. Sending each one to a large model is slow and expensive. Myles makes them in one pass, with a probability you can set a threshold on.
The problem
Guardrails add latency
Every safety and routing check that calls a large model adds hundreds of milliseconds and a bill to each turn.
Text answers are fragile
Parsing a yes or a label out of generated text breaks in production, and the confidence is guesswork.
Data leaves your product
Sending user content to another provider for every check is a privacy and compliance cost.
Where Myles sits
Why Myles
Structured by design
Yes or no, one of several labels, or a level on a scale: a probability for every option, no parsing.
Predictable cost
The Decision API costs £0.0015 per million input tokens from prepaid credit, or run it yourself under a licence with no per-decision fees.
The evidence
91.8%correct on 1,400 held-out decisions, ahead of the medium tier of a leading hosted LLM family
2.7%error on the 73% of decisions kept at a 5% target, on unseen test data
£4per million decisions on one CPU thread, against £1,880 for that hosted tier
Myles-400M, our own measurement, September 2026. Method and all results