DeepSeek V4 Pro Context Caching: Prices & How to Max Your Hits

DeepSeek's context caching is **automatic prefix caching**: if a later request starts with the same token prefix as an earlier one, the replayed prefix is billed at the cache-hit rate instead of the miss rate. No headers, no SDK changes, no configuration^[2][3].

The discount applies per token replayed, including partial prefix matches. In practice, any workload that reuses a large, stable prefix — a system prompt, a tool schema, a long document, a codebase summary — sees most of its input billed at the hit rate.

The same mechanism powered the preview-era pricing (cache hit was $0.003625 vs $0.435 miss, a 99.2% discount). The August 16 update kept the mechanism but changed the rates^[2].

• Automatic prefix caching — nothing to enable
• Hit rate applies to any replayed prefix, partial matches included
• No minimum prefix length documented by DeepSeek; in practice keep prefixes stable and long
• Works on both V4 Pro and V4 Flash

Cache-hit input is still by far the cheapest line on the V4 Pro price sheet — but the gap narrowed after the peak/off-peak update^[2].

Relative to the old flat rate, cache-hit input went from $0.003625 to $0.022 off-peak (+507%) and $0.044 peak (+1,114%) — the largest percentage increase in the whole price update, and the source of the '1100% price hike' headline. Even so, a 30x discount on input remains the single biggest cost lever DeepSeek offers^[2][7].

The pricing page explains the full old-vs-new comparison and the headline math.

A realistic agent loop shows the effect. Consider a 20-turn conversation with a 50K-token system prompt + tool schema^[2].

The takeaway: on multi-turn workloads, the prefix (system prompt + tools + history you keep stable) is where the money goes. Keeping it stable and replayable turns a 30x-discounted cost into a rounding error.

Even at the new, higher cache-hit rates, a 96% saving on input tokens dwarfs any other line-item optimization^[2].

# Off-peak rates, per 1M tokens
hit, miss = 0.022, 0.66

# 20 turns, 50K shared prefix replayed each turn after turn 1
prefix = 50_000
turns = 20
miss_cost = prefix * miss / 1e6          # turn 1 prefix: $0.033
hit_cost  = (turns - 1) * prefix * hit / 1e6  # turns 2-20: $0.0209
saved = (turns - 1) * prefix * (miss - hit) / 1e6  # ≈ $0.63 per 50K prefix

# At 1M tokens of shared prefix replayed 19 times:
#  cache-miss cost: $12.54   cache-hit cost: $0.42   → 96.6% cheaper

Cache hits are earned by prefix stability, so prompt design is a cost discipline^[2][3].

The same rules apply to the DeepSeek Harness: harness configurations that keep a stable instruction header and append per-turn state at the tail will see materially higher hit ratios — DeepSeek's own agent benchmark config (stable prompt, max effort) is the reference pattern^[4].

• Keep the system prompt byte-stable across turns — no timestamps, no per-turn injections at the front
• Put dynamic content (current time, user state) at the end of the context, not the start
• Freeze tool schemas; appending a tool later invalidates the cached prefix from that point
• For RAG, append the document set after the stable system+tools block
• Monitor cache-hit ratio in your telemetry; a sudden drop usually means a prompt change

Agent frameworks are the biggest cache winners and losers. A harness that re-injects a long instruction block every turn — or reorders tool schemas between calls — destroys prefix reuse^[2][4].

DeepSeek's own agent benchmarks run with a fixed prompt, `top_p=0.95`, temperature 1.0, and the Harness minimal mode — a deliberately stable configuration. Teams running harness-based agents on V4 Pro should copy that discipline: freeze the instruction header, append state, and keep tool definitions stable across the loop.

When you combine hour-aware scheduling (peak vs off-peak) with cache-maximizing prompts, the combined saving on a 24/7 agent fleet is the difference between a V4 Pro bill that looks like a flagship price and one that looks like a rounding error.