What this is
At midnight on August 17, DeepSeek's new pricing took effect. The biggest change isn't that things got "more expensive" — it's how they got expensive: a precision strike on the lifeline of high-frequency cache hit scenarios.
The specifics: V4-Pro peak hours (9 AM–12 PM and 2 PM–6 PM Beijing time) saw cache hits jump from ¥0.025 to ¥0.30 per million tokens — a 12x increase. V4-Flash cache hits in peak hours rose 5x; output pricing rose 2.25x or more across the board. The only modest increase was on input misses, at 1.5x.
What matters here: the cache hit tier was DeepSeek's most visible concession to high-frequency callers — RAG (retrieval-augmented generation, where each answer queries a knowledge base first), Agents (AI assistants that autonomously plan steps and loop through tool calls), and long-context reuse all relied on it to keep costs down. This price hike claws back most of that concession.
Industry view
The first wave of reaction came from the developer community. A technical writer benchmarked it: a RAG project that used cache hits to push per-Q&A cost down to a few fen, now sees peak-hour single-call cost rise to over 2 li each — "the gap shows up after tens of thousands of calls." These voices are uniformly pragmatic: split static prompt segments, move batch tasks to nighttime, tier models by task, compress output — no complaints, just engineering solutions.
Another camp was less optimistic. Practitioners noted the new pricing table doesn't specify concurrency limits (the old Flash was 2500, Pro was 500): "If the price hike and concurrency cap hit at the same time, daytime batch jobs take a double hit." A bigger bill is a small problem; throttled call capacity is the real risk.
One level deeper: this move signals that China's LLM API market has entered a "subsidy retreat" cycle. Competing on token unit price is no longer the core capability — competing on call structure, hit rate, and task tiering is what actually matters. For companies fully dependent on DeepSeek with no fallback, this is a forced architecture upgrade test.
Impact on regular people
For enterprise IT: If your company already runs high-frequency AI Q&A, knowledge retrieval, or intelligent customer service, the bill structure will change immediately. We recommend having engineering pull the past week's logs first — check cache hit ratio and peak-hour share — then decide whether to adjust scheduling.
For working professionals: Limited direct impact on non-technical white-collar workers. But if your workflow leans heavily on DeepSeek for long-document processing, auto-generation, or email drafting, we recommend tracking your monthly bill yourself — the per-call difference is small, but accumulated over time it may no longer be "free lunch."
For the consumer market: Consumers won't feel it directly — the price hike targets B2B APIs. But the signal is worth noting: when top-tier model companies dare to raise prices, we read it as compute tight and subsidies contracting. The marginal cost curve of AI services is quietly shifting.