DeepSeek V4: Flash and Pro
DeepSeek's current lineup is the V4 generation, shipped April 2026 in two variants. DeepSeek V4 Flash is priced at $0.14 input / $0.28 output per million tokens, the high-volume default, and one of the cheapest capable models across every provider covered on this site. DeepSeek V4 Pro is priced at $0.435 input / $0.87 output per million tokens, roughly three times Flash's rate, for workloads that need stronger reasoning or output quality. Both share a 1-million-token context window and a 384,000-token maximum output length per response, and both can run in either a non-thinking mode (a direct answer) or a thinking mode (extended reasoning before the final answer), the mode you pick changes output length and therefore cost, not the per-token rate itself.
DeepSeek's pricing position
DeepSeek has built its reputation partly on price: its API has consistently been priced well below equivalent-tier models from OpenAI, Anthropic, and Google, on both input and output tokens per million, and V4 continues that positioning rather than breaking from it. If cost per token is your dominant constraint and your task doesn't require a specific competitor's particular strengths, DeepSeek V4 is worth putting at the top of your comparison list, Flash first, Pro if Flash's quality falls short on your specific task.
Automatic cache-hit pricing: a genuine cost lever
DeepSeek automatically discounts input tokens whenever a call's prompt prefix matches something it processed recently. There's no setup, no opt-in flag, and no separate caching API to integrate. On V4 Flash, a cache hit cuts the input rate from $0.14 to $0.0028 per million tokens, about 50× cheaper. On V4 Pro, a cache hit cuts input from $0.435 to $0.003625, over 100× cheaper. This matters most for exactly the pattern most production traffic already has: a long, unchanging system prompt or reference document sent at the start of every call, followed by a short, changing user message. Keep the stable part first and byte-identical across calls, and DeepSeek's cache does the rest, the saving shows up automatically on your bill without you having to build anything.
Retiring the legacy names: deepseek-chat and deepseek-reasoner
If you or your tooling still calls the older model identifiers deepseek-chat or deepseek-reasoner, know that both are aliases pointing at V4 Flash today (deepseek-chat resolves to Flash's non-thinking mode, deepseek-reasoner to its thinking mode), and DeepSeek has set 2026-07-24 as the retirement date for both the aliases and the old V3.2 generation they used to point to. After that date, calls using the legacy names will stop working. If you're integrating fresh or maintaining an existing integration, switch to calling V4-Flash or V4-Pro by name directly rather than relying on the legacy aliases, so nothing breaks when the cutover happens.
How DeepSeek's pricing structure works
Like every provider on this site, DeepSeek bills input and output tokens separately, with output priced roughly double input on both V4 models, the same generation-is-more-expensive-than-reading pattern that holds across the whole market. What's distinctive about DeepSeek is simply how low both numbers sit relative to the field, and now, with automatic cache-hit pricing, how much further the effective input rate can drop on top of that for repeat-prefix traffic.
A worked example on DeepSeek V4 Flash
At V4 Flash's standard rate ($0.14 input / $0.28 output per million tokens), a 1,000-token prompt with a 500-token answer costs 1,000 ÷ 1,000,000 × $0.14 = $0.00014 for input, plus 500 ÷ 1,000,000 × $0.28 = $0.00014 for output, $0.00028 per call, or $28 a month at 100,000 calls. Now suppose 800 of those 1,000 input tokens are a stable system prompt that repeats identically on every call and hits DeepSeek's cache: those 800 tokens drop to the $0.0028 cache-hit rate, and only the remaining 200 fresh tokens bill at the full $0.14 rate, cutting the input side of the bill by roughly 80% before you've changed a single line of application logic.
Compare DeepSeek against the similarly budget-focused Mistral API pricing calculator and Llama API pricing calculator, or against flagship providers like OpenAI's GPT API pricing. For every provider side by side, use the all-provider hub calculator.
