DeepSeek retired the deepseek-chat and deepseek-reasoner API aliases on July 24 at 15:59 UTC. V4 Stable also shipped. Two migration traps: V4 Flash has thinking mode ON by default (changing prompt behavior and token costs), and deepseek-reasoner maps to V4 Flash, not V4 Pro. Teams that didn't prepare are now debugging broken pipelines.
DeepSeek's July 24 API deadline has arrived. At 15:59 UTC today, the deepseek-chat and deepseek-reasoner aliases stop working. If your pipeline still calls those names, it's breaking right now.
DeepSeek published the deprecation notice on their own pricing page months ago, but the migration trap has caught teams off guard. The old deepseek-chat pointed to DeepSeek V3. The replacement, deepseek-v4-flash, has thinking mode ON by default. That means your prompt behavior changes, your token consumption changes, and your latency profile changes, all at once, with no one-line swap. If you want the V4 equivalent of the old V3 chat experience, you need deepseek-v4-flash with thinking explicitly disabled.
The deepseek-reasoner alias has an even nastier gotcha. It maps to V4 Flash by default, not V4 Pro. If your application relied on deepseek-reasoner for high-quality reasoning output, you're about to get downgraded unless you switch to deepseek-v4-pro manually. That's the kind of detail that doesn't show up in a one-paragraph deprecation notice.
DeepSeek V4 stable also shipped today alongside the alias retirement. The V4 family now has three tiers: V4 Flash for budget workloads with thinking on by default, V4 Stable for general-purpose production use, and V4 Pro for maximum quality. The pricing differential is significant. V4 Flash runs at $0.28/$1.10 per million tokens. V4 Pro is roughly 4x that. Teams that were coasting on deepseek-chat at the old V3 price point are about to feel the difference.
This pattern is becoming familiar. Chinese AI labs launch aggressively cheap API access, build a developer ecosystem, then tighten the terms. DeepSeek ran a loss-leader API for months to establish mindshare. Now the free ride is over. The question is whether Western developers who adopted DeepSeek as a budget alternative will pay the new rates or bounce to the next subsidized Chinese API.
Kimi K3 open weights drop July 27. If DeepSeek's API migration annoys enough developers in the next 72 hours, Moonshot could pick up the refugees.
Get AI news in your inbox
Daily digest of what matters in AI.