DeepSeek V4 Flash vs V4 Pro: August 2026 Pricing Archive

This page preserves the August 21, 2026 V4 Flash and V4 Pro price snapshot. It is no longer the current buying guide. DeepSeek released V4.1 Flash on September 10 under the canonical deepseek-flash model ID; the old Flash aliases now route to it.

DeepSeek says deepseek-v4-pro will route to V4.1 Flash from September 14 at 04:00 UTC and use Flash pricing until V4.1 Pro launches. Read the current DeepSeek V4.1 Flash API guide for live prices, benchmarks, and migration steps.

DeepSeek currently exposes two V4 API model IDs: deepseek-v4-flash and deepseek-v4-pro. The Flash endpoint resolves to DeepSeek-V4-Flash-0731, while Pro resolves to DeepSeek-V4-Pro-0813. Both support non-thinking mode and thinking mode, with thinking enabled by default.

Both models provide a 1M context window and up to 384K output. DeepSeek offers an OpenAI-compatible base URL at https://api.deepseek.com and an Anthropic-compatible base URL at https://api.deepseek.com/anthropic. The comparison above reflects the official pricing page checked on August 21, 2026.

DeepSeek bills by total input and output token usage. Input is split into cache-hit and cache-miss rates, and each rate is further split into peak and idle windows. Idle rates are half of peak rates.

Fees are deducted from promotional balance before prepaid balance when both are available. DeepSeek states that product prices can change and advises users to recharge according to actual usage.

After the August 2026 price update, DeepSeek no longer publishes a single all-day rate. Peak hours are Beijing time 09:00–12:00 and 14:00–18:00. Every other hour, including nights and weekends outside those windows, is billed at the idle rate.

Schedule batch jobs, evaluations, and non-interactive backfills into idle windows when you can. Interactive production traffic that lands in peak hours should be budgeted at the higher rates, especially for cache-miss input and long thinking-mode output.

Both V4 models support JSON Output, Tool Calls, the Responses API, the Anthropic API format, beta chat-prefix completion, and beta FIM completion in non-thinking mode. Both can be called through OpenAI-compatible and Anthropic-compatible SDK configurations.

FIM completion remains limited to non-thinking mode. If you need fill-in-the-middle, disable thinking for that request rather than assuming the default thinking path will accept FIM.

A 1M context window allows both models to accept large documents, codebases, or long conversation histories in one request. The 384K maximum output is unusually large and can support extensive generation, but it should not be used as a default target.

Large contexts and outputs can raise latency and cost quickly, especially on cache misses during peak hours. Set explicit output limits, keep reusable prefixes stable for cache hits, and measure real token usage before scaling.

The official table lists concurrency limits of 2,500 for V4 Flash and 500 for V4 Pro. DeepSeek’s rate-limit documentation explains that upgraded API users can also be limited at the account level and per user_id.

Requests exceeding the per-user limit receive HTTP 429. Assign meaningful user_id values for workload isolation and backoff correctly on 429 responses instead of retrying immediately.

Start with V4 Flash when cost, concurrency, or high-volume processing matters. It has lower cache-miss and output rates and five times the listed concurrency of Pro. Both models now support the Responses API.

Evaluate V4 Pro for workloads where its higher-tier reasoning quality justifies a 3× cache-miss input rate and 3× output rate. Test both with representative prompts and include thinking-token output in the cost comparison.