AI models used across the platform

Summary#

Metric Vault runs on four AI providers, and the model is chosen per feature rather than per account. A summarizing widget runs on a small fast model; a long-form article or a strategic synthesis runs on a large one. This page lists every feature and the model behind it, so you can judge for yourself what is producing an answer.

Overview#

Four providers are in use.

Two supporting services also run models: Cloudflare Workers AI translates dynamic text, and an image provider generates blog artwork.

A rule runs through the whole table. Where a tool reports on real data, the model is given that data and asked to explain it, not to supply it. The extension's one-line summary, for example, is written from measured figures with an explicit instruction not to introduce any number that was not in the input. The model is a writer there, not a source. Where a tool measures what AI assistants say, no model is used to produce the answer at all: the assistants themselves are asked and their replies are read.

Model choices change as providers release new versions. This page is verified against the shipped code and re-checked when it moves.

Research and analysis#

Writing, synthesis and checking#

The assistants measured by AI visibility#

The AI visibility tools do not use a model to write an answer. They send a real category prompt to each assistant and read what comes back. Every available engine is asked at the same time, and an engine that does not respond is left out of the result rather than estimated.

Each probe is capped at a short response so that comparisons across engines are like for like. See Visibility Overview and Google AI Overview Tracker.

Supporting models#

Translation covers 20 languages in the dynamic path: English, Spanish, French, Portuguese, German, Russian, Italian, Dutch, Japanese, Korean, Chinese, Arabic, Hindi, Turkish, Polish, Swedish, Indonesian, Vietnamese, Thai and Ukrainian. The interface itself ships in six. See Supported languages.

Translation fails open: if every provider errors, the original English text is returned rather than a blank space.

Premium models on Agency and above#

The plan comparison lists Premium AI models (Claude Opus) as included from the Agency plan upward, alongside Enterprise and Custom. It is not included on Free or Pro. See Full plan comparison.

This is enforced. On Agency, Enterprise and Custom, the features listed below run on claude-opus-4-8. On Free, Starter and Pro they run on the standard model instead.

Nothing is blocked or refused. Every feature works on every plan that includes it; only the model behind it changes. You will not see an error, an upgrade prompt or a different screen, because the entitlement is about which model does the work, not about whether you may do it. The plan checked is the one belonging to the site owner, so an author or editor working on someone else's site gets that owner's model, which is also whose credits the run draws.

A small number of shorter passes always use the standard model on every plan, including Agency: post translation, and the keyword-overview entity pass. These are short, structured jobs where the premium model makes no practical difference to the output.

If the plan cannot be determined at the moment of a run, for example during a brief database problem, the standard model is used. That keeps the feature working rather than failing, and it is the reason an Agency account may very occasionally see a run that was not on the premium model.

What else is enforced automatically, and what you will run into first:

The features that run on claude-opus-4-8 for Agency and above are Growth Actions, and the Blog Studio writing, humanizing, title, category, excerpt, originality, fact-check and image-prompt passes. On lower plans those same features run on the standard model.

Frequently asked#

Is my content used to train these models? Metric Vault sends your prompts and the grounding data to the provider to get an answer back. What each provider receives is listed in Subprocessors and data flow.

Why does the Full Article Writer use a smaller model than the other writers? Speed. A long article on the larger model takes long enough that the request would exceed the platform's time budget and fail. A finished article beats a timeout.

Why do the AI visibility probes use small models? Because they are not doing the work, they are being measured. The question is what each assistant says about a brand, and the answer needs to be comparable across engines and cheap enough to run at the volume tracking requires.

What happens if a provider is down? It depends on the feature. The AI visibility tools drop the engine that did not answer and report the ones that did. The content checks try the providers in turn and only report AI as unavailable when all of them fail. Translation returns the original text. A tool that depends on one provider returns its error rather than a fabricated result.

Can an administrator change which model runs a feature? Several of the Claude-backed features read their model from platform configuration, so they can be moved to a different model without a code change. That is an operator action, not an account setting.