The four models
One engine, four selection policies. Every Lumen tier reads the same request, scores the same internal lanes, and picks a seat — what you choose is how far it will spend for quality.
One engine. Four points on the curve.
Same router, same lanes, same learning ledger, on every tier — a tier is a selection policy over that identical, already-scored slate. Read left to right: cost-optimized to quality-ceiling. On an easy request all four may still pick the same lane, and that's the design working.
- 1
Lumen Flash
Cheapest lane that clears the bar. Maximum thrift.
sylphx/lumen-flashBuilt forInteractive apps, chat surfaces, high-volume light workSelection policyTakes the cheapest lane that clears the quality bar, accepting a wide band below the predicted best.Escalation speedUpgrades its driving lane after four straight signs of session distress — the stickiest tier in the family, because thrift is what you bought. - 2
Lumen
DefaultThe default. Best value in the family.
sylphx/lumenBuilt forEveryday agentic workSelection policyTakes the cheapest lane near the predicted best, with a tight tolerance.Escalation speedUpgrades its driving lane after three signs of distress. - 3
Lumen Pro
The predicted best. You pay for the ceiling.
sylphx/lumen-proBuilt forComplex agentic work, planning, hard reasoningSelection policyTakes the predicted-best lane; cost only breaks near-ties.Escalation speedUpgrades its driving lane after two signs of distress — recovers from tool failures faster than Standard or Flash. - 4
Lumen Ultra
Council on demandZero tolerance — plus a council when it's hard.
sylphx/lumen-ultraBuilt forResearch, critique, the decisions that matterSelection policyTakes the predicted best with zero tolerance. When a request looks hard or unfamiliar, it convenes a multi-lane council and an independent judge picks the served answer.Escalation speedUpgrades its driving lane after a single sign of distress — the fastest reaction in the family.
Choosing a tier
Full capability — vision, tools, streaming, English and 中文 — is on every tier. You are choosing a quality ceiling, not a feature set.
- Latency-sensitive UI or bulk light tasksLumen Flash
- Not sure which tier you needLumen
- Your agent plans, architects, or reasons hardLumen Pro
- The answer is worth waiting for when it's hardLumen Ultra
Certified, not claimed
Tier order isn't asserted once and left alone — it's checked continuously against production evidence.
Flash ≤ Standard ≤ Pro ≤ Ultra in judged quality is verified from real duels on live traffic, every day. If two neighbors ever stop separating beyond noise, they merge — four is a hypothesis the data has to keep earning, not doctrine.
See the full certification ladder and daily trajectory