DeepSeek retires deepseek-chat and -reasoner Jul 24; move to V4
Anything pinned to the old model IDs breaks tomorrow. The V4 replacements are MIT open weights with 1M context — and a peak-hour pricing tax to route around.
Copy markdown
The 15:59 UTC cutoff
deepseek-chat and deepseek-reasoner are fully retired after Jul 24 at 15:59 UTC — code pinned to those model IDs starts erroring tomorrow. DeepSeek's migration path is the V4 endpoints, which accept both OpenAI- and Anthropic-format requests.
V4-Pro and V4-Flash, open under MIT
The replacements are a 1.6T-parameter MoE with 49B active (Pro) and a 284B/13B Flash, both open-weighted on Hugging Face under MIT with a 1M-token context window by default. You can self-host or fine-tune the exact models DeepSeek serves.
$0.87 per M out, cache hits near-free
Off-peak, V4-Pro runs about $0.44 in / $0.87 out per million tokens and Flash about $0.14 / $0.28. Cached input falls to roughly $0.004/M, so agent loops that resend the same context get much cheaper.
Your API bill now depends on the clock
V4 adds peak-valley pricing: rates double from 09:00–12:00 and 14:00–18:00 Beijing time. Those windows land overnight in US Eastern, so batch and cron jobs run off-peak effectively pay half.
Frontier-class coding, self-hosted
Early analyses put V4-Pro atop the open-weight SWE-bench Verified leaderboard near 80%, close to closed frontier models, while using a fraction of V3.2's FLOPs and KV cache. A near-frontier coder you can actually run yourself.