DeepSeek V4-Pro goes GA: native Codex support, open weights on HF
Native OpenAI Responses format means a one-script Codex swap; MIT weights, low/high/max reasoning, and today's 16:00 UTC price cutover round it out.
Copy markdown
Codex points at DeepSeek with one script
The API now speaks the OpenAI Responses format natively, and DeepSeek ships a one-line installer that wires Codex straight to `deepseek-v4-pro`. Native compatibility isn't identical behavior, so smoke-test your agent flow before trusting it in production.
MIT open weights, up on Hugging Face
The full DeepSeek-V4-Pro-0813 checkpoint — a 1.6T-parameter MoE (~49B active) with a 1M-token context — is live under an MIT license, so you can self-host or fine-tune with no strings. Caveat: the repo shipped without a Jinja chat template, so you supply the prompt format yourself.
Three reasoning tiers, and 'high' is the default
V4-Pro and V4-Flash now take low, high, or max thinking effort. Start at high and route only the hard tail up to max — max burns far more tokens for marginal gains on routine agent work.
Agent scores leap — on DeepSeek's own harness
Against the April preview, DeepSWE jumped 12.8 to 62.7 and CyberGym 52.7 to 83.3, with Terminal Bench 2.1 at 87.9. Every figure is DeepSeek's own and unverified externally, and its table still trails Kimi K3, Opus 4.8, and Fable 5 on several benchmarks.
New peak/off-peak pricing goes live 16:00 UTC today
The flagged cutover lands today: off-peak is half of peak, but both sit above the old flat rate. Batch work into off-peak windows to soften the hike — the `deepseek-v4-pro` alias auto-routes to the GA build, so timing is your only lever.