DeepSeek V4 Flash vs V4 Pro: Models & API Pricing

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. 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 10, 2026.

DeepSeek bills by total input and output token usage. Input is split into cache-hit and cache-miss rates, creating a large cost difference for applications that reuse stable prompt prefixes.

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.

DeepSeek’s official pricing page currently warns that API service pricing is expected to increase across the board in the near term and that the increase may be significant. No final replacement rates are listed on the page yet.

Treat every number in this guide as a dated snapshot rather than a future price guarantee. Recheck the official pricing page before purchasing a large balance, calculating a long-term budget, or launching a high-volume workload.

Both V4 models support JSON Output, Tool Calls, 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.

The official table still marks Responses API support as available only for deepseek-v4-flash and not yet available for deepseek-v4-pro. Although the page says Pro support was planned for early August 2026, its current status field has not yet changed, so this guide preserves the official status instead of assuming rollout completion.

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. 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, Responses API support, or high-volume processing matters. It has lower cache-miss and output rates and five times the listed concurrency of Pro.

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.