DeepSeek prices cached agent input at $0.003 a million tokens off-peak

An agent that keeps a 500,000-token prefix warm and reads it back across 100 requests consumes 50 million cached input tokens. In one simplified example published with the launch, that read bill comes to about $0.15 off-peak on V4.1-Flash. The same reads run roughly $15 on Moonshot's Kimi K3, $20 on GPT-5.6 Sol and $25 on Claude Opus 5, before cache writes, fresh context or output ^[8]. What decides the comparison is the cache-read share of your own bill. Those same reads cost $25 on Claude Opus 5, and Anthropic charges $25 per million output tokens, so the read bill in that example equals the price of one million output tokens ^[8][6]. Spread across the 100 requests, that is 10,000 output tokens per request ^[4]. An agent that writes more than that per turn is output-dominated. Then the comparison that decides the switch is $0.60 against $25, a factor of about 42, not $0.003 against $0.50, a factor of about 167 ^[3][1]. DeepSeek's own reasoning-effort setting pushes workloads that way. The company said raising reasoning effort from 25 to 100 improves results but can consume about 2.5 times as many output tokens ^[16]. DeepSeek ran the benchmarks itself, and the margin is narrow. On DeepSWE v1.1 it reported 74.2 for V4.1-Flash against 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol, two-tenths of a point over Anthropic on the vendor's own harness, plus 88.1 on CyberGym and 54.8 on AutomationBench ^[15][8]. The rollout risk is one DeepSeek names in the same release. The company said sparse-selection errors and approximate state reconstruction could degrade capability in edge cases, singling out sparse retrieval over very long contexts and cache-resumption boundaries ^[14]. Those are the operations a cache-heavy agent performs constantly, since the cost case rests on systems that reread the same repository, tool definitions, system instructions and conversation history ^[4]. Some buyers have no choice about migrating. Legacy V4 Flash requests are already served by V4.1-Flash, and DeepSeek's documentation says V4 Pro will route to V4.1-Flash after Sept. 14, 2026 until a future V4.1 Pro release ^[10]. The documentation says the new model has "comprehensively surpassed V4 Pro in performance, cost, speed, and total time" ^[11]. The architecture underneath is different. It runs 40 Transformer layers split into a 20-layer causal encoder and a 20-layer decoder, with 8 billion parameters active per token on input and 16 billion on generation, and FP4 KV caching that DeepSeek said cuts the global KV cache to 890 bytes per token, about a quarter of V4-Flash's ^[12][13]. The off-peak rate covers most of the week. Peak is seven hours a day on weekdays, 35 of the 168 hours in a week, or about 21% ^[3][5]. Where your team sits decides whether that matters. The later peak window covers 08:00 to noon at UTC+2, the working morning in western Europe, while both windows fall outside the business day at UTC-4 ^[6]. Most buyers cannot size any of this. In VentureBeat Pulse Research's July 2026 survey of 170 enterprises with more than 100 employees, 53% said they do not rigorously track AI compute cost and return on investment ^[17]. Twelve percent said they had not addressed inference-memory limits such as KV-cache capacity and 7% were not aware of the constraint, 19% between them ^[18][7]. Reuters reported this week that DeepSeek has tapped CITIC Securities as it prepares for a potential listing on Shanghai's STAR Market. A current fundraising could value the company at as much as 500 billion yuan, Reuters said ^[19]. Two questions settle the decision. Your logs answer both. First, what share of your input tokens are cache reads, and how many output tokens a turn actually emits. Second, whether correctness depends on retrieval across a very long context or on resuming from a cache boundary. Cache-heavy work that never touches those boundaries is where the off-peak rate saves real money.