DeepSeek launches V4-Pro and V4-Flash, retiring legacy API model names in July 2026

DeepSeek released V4-Pro and V4-Flash via its API on April 24, 2026, marking a new generation of its language model lineup. The update also sets a hard deprecation deadline: the long-standing deepseek-chat and deepseek-reasoner model names will stop working on July 24, 2026.

What's new

Two new models are now available through the DeepSeek API:

• DeepSeek-V4-Pro — the flagship model of the V4 generation, targeting maximum capability
• DeepSeek-V4-Flash — a faster, lighter variant optimized for lower latency and cost

Both are accessible through two API interfaces:

• OpenAI ChatCompletions format (drop-in compatible)
• Anthropic Messages API format

The base URL is unchanged. Migrating requires only a model parameter update: set model to deepseek-v4-pro or deepseek-v4-flash.

Legacy model deprecation: The previous model names deepseek-chat and deepseek-reasoner currently map to V4-Flash's non-thinking and thinking modes respectively. Per the release notes: "The two legacy API model names, deepseek-chat and deepseek-reasoner, will be discontinued in three months (2026-07-24)." Developers should update integrations before that date.

Context

DeepSeek's V3 generation (December 2024) attracted widespread attention for delivering near-frontier benchmark performance at a fraction of the cost of OpenAI and Anthropic models. The V3 series included a Mixture-of-Experts architecture that reduced inference cost significantly while maintaining competitive reasoning quality.

V4 extends that lineage. The dual API interface support — OpenAI-compatible and Anthropic-compatible — continues DeepSeek's strategy of minimizing migration friction for developers already building on either major SDK. V4-Flash absorbs the role previously played by deepseek-reasoner as the thinking-mode entry point.

DeepSeek models are available directly through the DeepSeek API as well as third-party inference providers including Together AI, Fireworks, and others.

Why it matters

DeepSeek occupies a distinctive position in the AI landscape: a lab offering frontier-class models at substantially lower cost than western counterparts, under broadly permissive licensing. The V4 launch continues that pattern with a cleaner, two-model lineup (Pro and Flash) replacing the previous naming scheme.

The July 24 deprecation deadline is operationally significant. Any production system with deepseek-chat or deepseek-reasoner hardcoded as model strings will break without updates. The 90-day window is reasonable, but given how widely these model names are used in third-party tooling and tutorials, the transition will require active coordination across the developer ecosystem.