DeepSeek retires legacy models July 24, launches V4 with new pricing

DeepSeek has officially announced the retirement of its legacy API model names, deepseek-chat and deepseek-reasoner, effective July 24, 2026 at 15:59 UTC. After that time, any request using these names will return an error with no fallback. The V4 preview, launched April 24, 2026, introduced two replacement models: deepseek-v4-flash (284B parameters, 13B active) and deepseek-v4-pro (1.6T parameters). Flash outputs cost $0.28 per 1M tokens, Pro outputs $0.87 per 1M tokens. Both support a 1M-token context window and 384K maximum output. A critical trap exists: deepseek-reasoner currently points to V4-flash's thinking mode, not V4-pro. Teams assuming "reasoner" means top-tier reasoning will experience a silent quality downgrade if they naively rename to deepseek-v4-flash. Only deliberately switching to deepseek-v4-pro preserves the high-reasoning capability.

Under the legacy system, reasoning was model-specific: choose deepseek-reasoner for chain-of-thought. With V4, thinking mode is now a request parameter, decoupling the reasoning capability from the model name. This simplifies routing but requires code changes. The migration itself is straightforward for most callers: change the model string in API requests, keep the base URL and API key unchanged. DeepSeek provides no automatic forward or soft-fail after July 24. Teams must also re-evaluate which workloads need Pro vs. Flash reasoning quality to avoid either overspending or losing accuracy. The deprecation date has been published since April 24 in DeepSeek's changelog, pricing page, and V4 release notes, giving developers a clear three-month window to update production code, test the new parameters, and adjust cost assumptions.