Skip to content
NewsProduct launch

DeepSeek V4-Pro reaches GA as deepseek-chat and deepseek-reasoner endpoints face July 24 retirement

· by Pondero Newsdesk

The short version

The deepseek-chat and deepseek-reasoner API endpoints retire July 24 at 15:59 UTC. Developers who do not update their model string before that cutoff will hit hard 404 errors.

DeepSeek V4-Pro reaches GA as deepseek-chat and deepseek-reasoner endpoints face July 24 retirement

Eight days remain before DeepSeek shuts off the deepseek-chat and deepseek-reasoner endpoints. Any application still calling those model strings after July 24 at 15:59 UTC will receive a hard 404, per DeepSeek's API changelog. There is no grace period indicated in the docs.

What

DeepSeek launched V4-Pro and V4-Flash as a preview on April 24, 2026. Both models are now in general availability, per TechNode's June 30 reporting. The endpoint retirement notice has been in the official changelog since the April launch: "deepseek-chat and deepseek-reasoner will be fully retired and inaccessible after Jul 24th, 2026, 15:59 (UTC Time)."

The migration path requires one code change. DeepSeek's changelog says to keep the same base URL and update the model string to either deepseek-v4-flash or deepseek-v4-pro. Right now, deepseek-chat routes internally to deepseek-v4-flash in non-thinking mode. deepseek-reasoner routes to deepseek-v4-flash in thinking mode. That silent compatibility layer disappears July 24.

V4-Pro carries 1.6 trillion total parameters with 49 billion active per forward pass. V4-Flash carries 284 billion total with 13 billion active. Both support a 1M-token context window and both operate in standard (non-thinking) and extended reasoning (thinking) modes, per the official release notes. Both support the OpenAI ChatCompletions API format and the Anthropic API format.

DeepSeek's pricing page lists V4-Flash at $0.14 per million input tokens (cache miss) and $0.28 per million output tokens. V4-Pro runs $0.435 per million input tokens (cache miss) and $0.87 per million output tokens. Cache hits drop to $0.0028 and $0.003625 respectively. TechNode reported in June that DeepSeek planned a 2x peak-hour surcharge during Beijing business hours (9 AM to 6 PM CST, 1 AM to 10 AM UTC); that multiplier does not appear on the current pricing page.

Why it matters

The old endpoints abstracted model selection. Teams building on deepseek-chat or deepseek-reasoner never chose between Flash and Pro. The July 24 cutoff forces that decision: V4-Pro output tokens cost three times more than V4-Flash, so the choice has real per-call cost consequences.

The failure mode in production is also unforgiving. A 404 from a model API typically surfaces as an unhandled exception in agentic pipelines built on n8n, Make, Claude Code, or similar orchestration layers. Most pipeline configs do not gracefully degrade on model-not-found errors. That makes a deadline-day migration riskier than a staged one done now, with time to test and observe costs.

There is a secondary nuance worth flagging: teams that assumed deepseek-reasoner gave them V4-Pro-grade reasoning were actually getting V4-Flash in thinking mode. V4-Pro is a distinct, larger model with separate pricing. Switching to deepseek-v4-pro explicitly will produce different behavior and a higher bill.

What to watch next

An official GA announcement on deepseek.com had not appeared as of July 16. Independent benchmarks comparing V4-Pro against GPT-5.6 Sol and Claude Sonnet 5 on cost-per-token efficiency are expected from third-party researchers. Teams with tight budgets should verify the current pricing page directly before locking in Q3 API spend, since DeepSeek's pricing page notes that "product prices may vary and DeepSeek reserves the right to adjust them."

Sources