WordPress AI in core keeps slipping. The plugin does not

WordPress 7.1 "Mary Lou" shipped on August 19, 2026, on the closing day of WordCamp US in Phoenix. Its own roadmap, published two months earlier, promised four AI items. Read the field guide and most of them are not there.

That is not a scandal. It is a planning signal, and it is worth reading before you quote AI work on a client site. The pattern across three releases is consistent: the AI infrastructure lands in core, and the AI features land in a plugin. This post lays out what actually shipped in each release, what moved, and what that means for the thing you are about to build.

What did WordPress 7.1 actually ship for AI?

Abilities API work, and very little else with an AI label on it. The Abilities API is the core registry where plugins, themes and core describe what they can do in a machine-readable format, so that automation tools and AI agents can call those capabilities predictably. In 7.1 it gained filtering, execution lifecycle hooks, JSON Schema preparation and a unified public exposure flag (WordPress 7.1 field guide).

The headline features of 7.1 are not AI features at all. The release brings a Playlist block, Notes with suggestion mode and emoji reactions, and client-side image resizing that moves the work into the browser before an upload leaves the machine (WordPress 7.1 Release Candidate 1). Anne McCarthy led the release, which is named for the jazz pianist Mary Lou Williams.

So the honest summary is narrow. WordPress 7.1 improved the plumbing that AI tools call, and shipped no new AI capability that a site owner can see.

What was promised, and what moved?

The Roadmap to 7.1 named four AI workstreams on June 19, 2026 (Roadmap to 7.1). Here is each one against what the field guide and the AI team reported afterwards:

Promised in the 7.1 roadmap	Status after release
AI Client generation streaming	Not documented in the 7.1 field guide
Embeddings for meaning-based search	Missed Beta 1; targeted at WordPress 7.2
Connector authentication beyond API keys	Not documented in the 7.1 field guide
Guidelines, with import and export between sites	Not documented in the 7.1 field guide
Abilities API expansion	Shipped: filtering, lifecycle hooks, schema prep, unified public flag

Embeddings are numeric representations of content that let a site search by meaning rather than by matching words. The roadmap carried its own disclaimer, and it is a fair one: what is shared there is being actively pursued, but does not necessarily reach the final release. The AI team's contributor summary of August 20, 2026 confirmed the embeddings deferral directly, saying the integration did not make the Beta 1 cutoff and is targeted for 7.2, to allow more testing across multiple providers (Make WordPress AI).

One more scheduling note matters for anyone planning a 2026 build. React 19 was punted past 7.1, so WordPress 7.1 stays on React 18.3 (Developer Blog, August 2026).

Where is the AI actually shipping?

In the canonical AI plugin, on a roughly monthly cadence. That is the part the release-cycle coverage keeps missing. The canonical AI plugin is the official plugin that carries WordPress AI features outside the core release schedule (AI team release posts).

Version 1.2.0 arrived on July 21, 2026 with a comment-moderation reply experiment, bulk content summaries across multiple posts, and two new abilities, core/read-content and core/read-users. Version 1.3.0 followed on August 18, 2026 with a content translation experiment, slug generation, an opt-in setting for custom abilities, settings import and export, and better content classification.

Set the two tracks side by side and the division of labour is obvious:

Release track	What it delivers	Cadence
WordPress core	Registries and APIs: Abilities API, AI Client, connectors	Two or three releases a year
Canonical AI plugin	Features users click: alt text, classification, translation, slugs	Roughly monthly

This is a deliberate design. Core does not bundle AI providers either; they are shipped as separate plugins so they can iterate faster than the core cycle allows. The same logic explains why the features sit in a plugin.

What does this mean for a build I am quoting now?

Build against the Abilities API today, and treat every user-facing AI feature as a plugin dependency rather than a core guarantee. The AI Client is the provider-agnostic PHP API that core added in 7.0 so plugins can send prompts without knowing which vendor answers (AI Client announcement).

Concretely:

1. Registering abilities is safe. The API is in core since 6.9 and gained a real expansion in 7.1. It is the most stable AI-adjacent surface WordPress has.
2. Calling models through the AI Client is safe. It landed in 7.0 on May 20, 2026, and wp_ai_client_prompt() is the documented entry point.
3. Do not promise semantic search from core. Embeddings are 7.2 material at the earliest, and 7.2 has no date.
4. Do not promise a Guidelines feature. It appeared on a roadmap and not in a field guide.
5. Price plugin features as plugin features. The canonical plugin's translation and classification work is labelled experimental, and experiments change.

Consider a straightforward client ask: "can the site suggest tags for new posts?" The honest answer today is yes, via the canonical plugin's classification feature, not via core. That distinction changes who supports it in a year.

Why does core AI keep slipping?

Because the AI work is dependency-shaped, and dependencies slip in a way that features do not. For example, core supports several providers at once, so anything model-facing has to be proven against all of them (AI contributor summary, August 20, 2026). Embeddings are the clearest case: the team held them back specifically to test across multiple AI providers, since core supports no single provider of its own.

There is a structural reason too. Anything core ships becomes a compatibility promise for years, across every host and every provider plugin. A translation experiment in a canonical plugin can be revised in four weeks. A registry API in core cannot. Slipping the risky half into a plugin is a reasonable trade rather than a failure, even when it makes release announcements read thinner than the roadmap did.

The pattern also rewards attention. Anyone tracking WordPress AI by release notes alone will conclude the project is slow. Anyone tracking the AI plugin's monthly releases sees the opposite. Judging tooling by shipping cadence rather than announcements is the same discipline behind ranking any AI agent framework 2026 by churn instead of by feature lists.

Read the field guide, not the roadmap

The gap between the two documents is the whole story here. A roadmap is a statement of intent published two months out. A field guide is a description of what shipped, and for WordPress AI in 2026 the two keep diverging.

None of that makes core AI a bad bet. It makes it a slow, deliberate one, with the fast-moving work parked in a plugin where it can be revised. If you are choosing what to build on this quarter, take the registry from core, take the features from the plugin, and keep the two clearly separated in whatever you promise a client.