DeepSeek's price change hits Sunday 16:00 UTC. I ran 650 real API calls through both price tables — the more you benefit from caching today, the harder your bill jumps (mine: 2.7x, not the 1.5x peo...

Everyone's focused on the output price going up. If you use agents or long conversations, that's not the tier that matters. I pulled every request out of my local session logs: 650 requests, 16 sessions, 155.9M prompt tokens, 98.09% cache hit rate (normal for agent workloads — the same context gets resent constantly, and almost all of it hits cache). Repriced those exact tokens under the new table: my bill comes out at 2.7x . The same tokens with zero cache hits would only be 1.5x. In other words: the better caching works for you today, the bigger your increase on Sunday. Here's why — the three tiers didn't move together (v4-pro, off-peak, per 1M tokens): tier old flat price new off-peak change input, cache hit $0.003625 $0.022 6.1x input, cache miss $0.435 $0.66 1.5x output $0.87 $1.98 2.3x The cheapest tier is the one that moved 6x. Caching is still absolutely worth it (on my load it still cuts the bill ~15x vs no cache) — it's just a much less generous discount than before. One more thing most posts missed: the new pricing has peak and off-peak windows, and off-peak is half price. Peak is 01:00–04:00 and 06:00–10:00 UTC. If you're in the US, your entire workday falls in off-peak on both coasts. If you're in China, peak lands exactly on working hours. Before anyone asks whether these numbers are real: I reconciled them against my DeepSeek console billing for the same period — 2% delta, explained by 34 requests missing from local logs (a deleted session plus failed retries). So everything above is a lower bound, not an estimate. I put the audit script and the full methodology online — links in the comments to keep this post clean. The script is a single Python file, standard library only, no network calls, never reads credentials.