OpenRouter's DeepSeek "Latest" Aliases Already Repriced: Flash Cache Reads 5× Off-Peak

One day after OpenRouter quietly launched the ~deepseek/deepseek-pro-latest and ~deepseek/deepseek-flash-latest rolling aliases, their pricing records in the public models API have already mutated — no changelog entry, no notice. The alias IDs are proving to be moving targets on price, not just on model pointer.

What changed

The Pro alias dropped its hour-of-day overrides. Yesterday's API record showed a 1.93× peak/off-peak split ($1.122/M input 00:00–14:00 UTC vs $0.5808/M after). Today it's a flat $0.96/M input, $2.88/M output — with cache read at $0.088/M. That's 2.68× the underlying pinned model's cache price: deepseek/deepseek-v4-pro-0813 charges $0.0328/M for cached input.

Flash is worse. The alias now lists $0.15/M input, $0.60/M output — but cache read is $0.015/M, versus $0.006/M peak and $0.003/M off-peak on the pinned deepseek-v4.1-flash. That's 2.5× peak, 5× off-peak, on the single line item agentic workloads hammer hardest. Long-context loops that re-read a 500K-token context every turn pay the alias premium on every cached token.

Why a builder cares

Cache-heavy is the new default: agent scaffolds, RAG pipelines, and code assistants live on cache reads, and cache is exactly where the alias premium concentrates. Same model, same tokens, 5× the cache line because you used the friendlier slug. Pin deepseek-v4.1-flash-20260910 or deepseek-v4-pro-20260813 and the discrepancy disappears. If an alias reprices within 24 hours of launch, treat every *-latest ID as an unaudited billing surface — check the pinned slug's pricing before each deploy.