DeepSeek V4 Flash vs Pro: the price gap, worked out.

Quick answer: Pro costs 3.1x, and most tokens do not need it.

At DeepSeek's published rates as of July 2026, V4 Pro bills 3.1x Flash on input ($0.435 vs $0.14 per million) and 3.1x on output ($0.87 vs $0.28). Both variants share the 1M context, the 384K output ceiling, and the thinking modes; the multiplier buys network scale (1.6T total parameters against 284B, by vendor figures) and the reasoning depth that comes with it. The decision is not which model is better, Pro is, but which tokens deserve the multiplier. On most production mixes the answer is a minority: chat, extraction, summarization, and routine agent steps hold up on Flash, and the workloads that genuinely need Pro announce themselves by failing on Flash in ways you can name. Route by workload and the blended rate lands much closer to Flash than to Pro.

The published rates, including cache pricing.

DeepSeek's API docs publish three prices per variant, and the cache-hit column is the one most comparisons skip. On the vendor's own platform, input tokens that hit the automatic context cache bill at $0.0028 per million on Flash and $0.003625 on Pro, a 50x and 120x reduction against the respective cache-miss rates. Cache behavior is covered in depth on the caching page linked below; the headline is that repeated prompt prefixes, system prompts, chat history, few-shot blocks, are where DeepSeek's platform quietly discounts agents and chat apps. Through APIsRouter, the catalog bills the flat input and output rates shown in the pricing table further down, without a separate cache-hit rate for the DeepSeek ids, so price plans conservatively at the flat rates and treat cache-hit savings as direct-platform behavior.

Worked math: four workload shapes on each variant.

The table prices four realistic workload shapes at the published cache-miss rates, the conservative case. Token counts are illustrative; the usage block on every response gives your real numbers, and the pattern to notice is how the absolute gap stays small on short-output work and widens where output volume grows.

When Pro earns its 3x.

The trap to avoid is defaulting the whole application to Pro out of caution. At a 3.1x multiplier the caution costs exactly 3.1x, compounds monthly, and mostly upgrades tokens that were fine. The blended rate of a routed mix, Pro on the named hard steps, Flash on everything else, is where this family's economics actually live.

• Multi-step reasoning that Flash demonstrably loses: plans that drift, chains that drop a constraint mid-sequence, tangled instructions that come back half-followed.
• Hard code: gnarly refactors, concurrency bugs, architecture-level review. Routine generation and test scaffolding stay on Flash.
• Long-document analysis where a wrong synthesis is expensive to catch: legal, financial, research material where review time costs more than the token multiplier.
• Planner seats in agent loops: one Pro call that plans well can save a dozen Flash calls that execute a bad plan, which makes the planner the single best place to spend the multiplier.
• The test stays the same: run the worst real prompt on Flash five times. Five passes means keep the lower rate; a nameable failure means that workload, and only that workload, moves to Pro.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Thinking mode: the output-side variable on both variants.

Both V4 variants support thinking and non-thinking modes, and reasoning tokens bill as output. That single fact moves the math more than the Flash-or-Pro choice on some workloads: a hard prompt that triggers a long reasoning pass can multiply completion_tokens well past the visible answer, on either variant, at that variant's output rate. Two habits keep it priced in. Log completion_tokens per request from day one, because averages over a mixed workload hide a tail of expensive reasoning-heavy calls. And set max_tokens with headroom on reasoning work: a truncated thinking pass returns finish_reason=length with little visible content, bills in full, and then bills again on the retry. The same discipline covers you on the GLM family, which shares the reasoning-billing shape at higher per-token rates.

How this fits the wider DeepSeek pricing picture.

This page stays deliberately narrow: the two V4 variants against each other. For the family against the rest of the market, the general DeepSeek pricing page linked below carries the cross-catalog tables; the one-line summary is that even Pro, the premium variant, prices under every Western flagship by a wide margin, so the Flash-vs-Pro question is about spending well inside what is already the market's lowest-priced capable family. For high-volume mixes it is also worth benchmarking outside the family entirely: glm-5 competes near Pro's band with a different reasoning profile, and MiniMax M2.7 sits in the same value tier. All of them are one model string away on the same endpoint, which is what makes these comparisons an afternoon instead of a procurement cycle. A useful sanity check on any proposed routing split is the blended rate. An 80/20 Flash-to-Pro mix at published rates works out to about $0.199 per million blended input and $0.398 blended output, roughly 40% above pure Flash and under half of pure Pro. If a proposed split blends much closer to Pro's rate than that, the escalation gate is letting routine tokens through, and the usage log will show which workload is leaking.

Who works this math.

• Agent builders splitting planner and executor seats, where the planner justifies Pro and the executor volume decides the bill.
• Chat and companion products, where per-turn cost times daily turns is the whole unit-economics story and Flash's rate sets the floor.
• Batch pipelines, classification, labeling, summarization at scale, where a 3.1x multiplier across millions of items is the difference between a rounding error and a budget line.
• Teams doing structured escalation: Flash by default, Pro behind a named-failure gate, reviewed monthly against the usage log.

FAQ