Concepts
Sessions, fleets, tiers, and the always-frontier baseline.
Session
A connection to one or more OpenAI-compatible hosts (BYOK). The raw vendor key never lands in localStorage. Pass your pmz_live_ key as Authorization: Bearer, or a short-lived session id as X-Promptimizer-Session.
Fleet
| Field | Meaning |
|---|---|
| id | Provider model id |
| tier | economy, standard, or frontier |
| input_per_1m / output_per_1m | USD, when known |
| source | catalog, heuristic, price, or user |
| selected | Eligible for routing |
Tiers
+−Economy
Mini, nano, haiku, 8B, instant. L1–L2 factual and short transforms.
+−Standard
Flash, sonnet, 70B. L3 code, analysis, medium reasoning.
+−Frontier
GPT-4 class, opus, o-series. L4–L5, plus high quality risk (system design, legal/medical, proofs).
Baseline
baseline_model is the expensive comparison. Saved cost uses the same token counts on both models. We do not invent a cheaper token count.
Quality risk
Independent of complexity. system_design, safety_sensitive, code_debug, and reasoning are high risk and cannot route to economy even if the prompt is short.