Everyone speaks Responses now
A competitor implemented a rival's API natively. That is not flattery — it is the moment an interface stops being a moat and becomes plumbing, with one hazard nobody mentions.
V4-Pro's general availability includes native support for the OpenAI Responses API, aimed at Codex-style setups.
Nobody implements a rival's interface to be polite
They implement it because the tools their customers already use speak that dialect, and the cost of not speaking it exceeds the cost of adopting someone else's shape. That is the definition of a standard, arrived at by capitulation rather than committee.
This concedes more than a completions shim
Every provider serves chat completions now, and that endpoint is thin. The Responses API is a richer surface — tool calls, state, structured output — so implementing it natively means adopting a rival's architecture, not just its URL.
Good news: providers become swappable. Hazard: a shared interface hides real behavioural differences and will not warn you.
The hazard, specifically
The same request to two models that both accept it can produce very different tool-calling behaviour — different willingness to call, different argument construction, different recovery from a failed call. The interface validates and passes it through. Nothing tells you the semantics moved.
Which means the abstraction layer everyone is now building against reduces integration cost and does nothing for evaluation cost. You still have to test every model against your own traffic, and the API's uniformity makes it feel like you do not.
Convergence on the surface, divergence underneath
Interfaces are standardising. Endpoints are collapsing into general models. And licences are diverging in the opposite direction.
Technically these models are becoming interchangeable. Legally and behaviourally they are becoming less so. Whichever of those you plan around will determine whether your abstraction layer is an asset or a comfort.
Digital Applied — AI Model Releases: August 2026 Tracker and Dated Ledger → · How AI Works — Best Open-Weight LLM for Agentic Coding in 2026 → · Mem0 — State of AI Agent Memory 2026 →