DeepSeek V4 Pro hits GA speaking the OpenAI Responses API
DeepSeek V4 Pro went GA with native OpenAI Responses API support and a reasoning-effort dial. Swapping models is now a base URL, not a rewrite.
DeepSeek moved V4 Pro to general availability today, ending a preview that started with the model's April debut. The spec sheet matters less than one line in the announcement: native support for OpenAI's Responses API, tuned for Codex. A Chinese lab shipping a frontier model that speaks a competitor's agent protocol is the portability story of the week — your harness stops being the thing that locks you in.
What actually happened
Per DeepSeek's own release notes and TechNode's writeup:
deepseek-v4-pronow resolves to DeepSeek-V4-Pro-0813. The model name doesn't change, so nothing in your code has to.- Native OpenAI Responses API support, with optimization called out for Codex, plus tool calls.
- A reasoning-effort dial across V4 Pro and V4 Flash: low for simple tasks, high for everyday agent workflows, max for the hard ones.
- 1M-token context, 384K max output.
- Current API pricing for V4 Pro: $0.003625 per million cached input tokens, $0.435 uncached input, $0.87 output.
- New pricing takes effect 16:00 UTC on August 16, 2026, with off-peak rates at half of peak.
DeepSeek's announcement includes a benchmark table. We're not repeating vendor-run numbers as fact — we said the same about Grok 4.6's scorecard. Run your own workload.
Why Responses API support matters for your stack
Model portability has always been sold as an API-shape problem, and for chat completions it basically was — swap a base URL and a key, done. Agents broke that. Once your automation depends on multi-turn tool calls, server-side state, and a specific tool-result envelope, you're not calling a model anymore, you're building against a runtime. That's where the lock-in moved.
So when a second lab implements the same agent protocol, the thing that gets cheaper isn't tokens — it's switching. If your agent already talks Responses API, DeepSeek V4 Pro becomes a base URL and a model string, and you can benchmark it against your incumbent on your own tasks in an afternoon instead of a sprint.
The reasoning-effort dial is the other lever, and it's the one most teams leave untouched. Most steps in a real agent loop are clerical: parse this, pick a tool, format that. Paying max-effort reasoning for a field extraction is the same mistake as paying an engineer to file receipts. Set effort per step, not per app, and the cheap steps get cheap.
One caveat we'd flag for anyone in a regulated or contract-sensitive spot: protocol compatibility is not the same as jurisdictional equivalence. Where the tokens are processed still matters for what you're allowed to send. Build the swap so it's possible, then decide per workload whether to use it.
Key takeaways
- DeepSeek V4 Pro is GA as DeepSeek-V4-Pro-0813 — same
deepseek-v4-promodel name, no code change - Native OpenAI Responses API support plus tool calls means switching agent backends is a base URL, not a rewrite
- Reasoning effort is now a per-request dial (low / high / max) — stop paying max for clerical steps
- 1M context, 384K max output, at $0.435 uncached input and $0.87 output per million tokens
- New peak/off-peak pricing lands 16:00 UTC August 16 — off-peak at half of peak; schedule batch work accordingly
- Benchmarks in the announcement are vendor-run; test against your own tasks before you switch
Is your agent hard-wired to one provider's runtime? We build agent stacks on the standard protocols so you can benchmark a new model on your real workload in a day and switch when the math changes. See how we build portable AI systems.
Sources: DeepSeek API release notes, DeepSeek API pricing, TechNode.
- #deepseek
- #responses-api
- #model-portability
- #ai-agents
- #llm-pricing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Zed Delta puts agent code review in one shared thread
Zed's Delta syncs the worktree and the agent conversation live across a team, with comments that stay anchored as code moves. Git still works underneath.
Read itA 16-year-old SQLite bug ate Tailscale's data
Tailscale hit 19 database corruptions in six months. The cause was a SQLite WAL race from 2010, triggered because they used SQLite in a non-standard way.
Read it