DeepSeek ships V4-Pro speaking OpenAI's agent API

DeepSeek moved V4-Pro to general availability this week with a feature that says as much about the protocol race as the model race: native support for OpenAI's Responses API, optimized for the Codex agent stack with one-click setup. The Chinese lab announced the release on Wednesday, billing it as a major agent upgrade with strong production gains.

The interop choice is the story. The Responses API is OpenAI's interface for tool use, state, and multi-step agent execution, and OpenAI has spent the year positioning it as the substrate for multi-agent applications. When a rival frontier lab implements that interface natively, it stops being one vendor's API and starts functioning as a de facto standard, the way S3's interface outgrew Amazon. For teams running Codex-style agent harnesses, DeepSeek is now a drop-in backend: same calls, same agent loop, different model. Model Context Protocol standardized how agents reach tools. The Responses API is quietly becoming the analogous layer for how harnesses reach models.

The GA release also formalizes tiered reasoning. V4-Pro and V4-Flash expose low, high, and max reasoning-effort settings, letting operators dial cost against depth per task, low for routine steps, high for daily agent workflows, max for hard problems. DeepSeek paired the release with a shift to time-of-day API pricing that rewards operators who can schedule batch agent work into cheaper hours, a change with its own implications for fleet economics.

For enterprise buyers, a native second source behind a shared interface changes procurement more than any benchmark. An agent harness written against Responses calls can now route the same loop to a second frontier lab without rewriting orchestration code, which means evaluations, failovers, and price negotiations all get cheaper to run. Model gateways and routers gain the most: when the interface is uniform, model choice collapses into a runtime decision, made per task on cost, latency, or capability, rather than an architecture commitment made once a year.

There is a governance asymmetry underneath the convenience. MCP and the A2A protocol now sit under foundation stewardship with published deprecation policies. The Responses API remains one company's product surface. OpenAI can evolve it in whatever direction advantages its own stack, and every third-party implementer, DeepSeek now included, inherits those changes on OpenAI's schedule. Standards bodies spent a year building neutral plumbing for agent interoperability; the market may be settling on a proprietary interface anyway, because it is the one the most popular harnesses already speak.

That tension, standardized interfaces on top, competing models and unsettled governance underneath, is the protocol story of the second half of 2026, and this release is the clearest data point yet.