Google Gemini
Has a free tier. This is what a new install starts on. On the free tier Google may use what you send to improve its products; the paid tier does not.
The five
NVIDIA NIM
Free credits, which NVIDIA describes as being for development. Check their terms before a production site depends on it.
OpenRouter
Has free models. OpenRouter says these carry low daily limits and are usually not suitable for production use.
Anthropic and OpenAI
No free tier. Both need a funded account billed by usage — and that is separate from any Claude or ChatGPT subscription you already pay for. A subscription does not include API access.
Keeping it working
Model lists refresh daily
Each provider's catalogue is fetched from that provider once a day, so the dropdown shows what your account can actually use — including models added since the last plugin release.
A withdrawn model is replaced
With the closest match from the same provider, and you are emailed. If nothing is close enough, nothing changes and the email says so — an unrelated model quietly writing your articles is worse than generation stopping.
Your provider is never changed
That is a key, a bill and a contract, not a detail to swap because a model moved.
Failover, in order
Set up to five model combinations. A failure moves to the next rather than losing the article, and a dashboard notice tells you when your first choice is the one failing.
How failover behaves once it fires
The fallback order is not a list of equals tried at random. It is an ordered chain that presents itself to the rest of the plugin as a single provider, so the code that writes an article does not know or care that failover exists. Two of its behaviours are worth understanding before you configure it.
It sticks
Once a call succeeds on a later entry in the chain, every remaining call for that article starts from there rather than returning to your first choice. This is deliberate. An article is not one request: the metadata, the body and the humanising pass are separate calls, and switching models between them produces voice drift you can hear when you read the result. A consistent second choice reads better than a mixture of a first and a second.
It is capped
At most three entries are tried per call by default, and five is the hard maximum however many combinations you configure. During a provider-wide outage an uncapped chain would work through every model you own a key for and still fail, just more slowly and at greater cost. The cap makes the failure arrive while it is still cheap.
Failures are classified, not guessed
Whether a failure is worth retrying somewhere else is decided from what the error actually was, rather than inferred from a status code. A malformed request is not an outage, and moving it down the chain would produce the same error four more times.
What a failed attempt costs you
A provider that accepts a request, consumes tokens and then fails has still billed you for those tokens. The usage figures for an article therefore include every provider the chain actually called, not only the one that returned the finished text.
This is the honest number rather than the flattering one. It is also the reason the attempt cap matters: each additional entry tried is a real charge on a real account, and the point of stopping at three is that the fourth and fifth rarely change the outcome.
Keeping keys out of the database
Any provider key can be defined in wp-config.php instead of stored in the database. The settings screen tells you the constant name for each provider. A key set that way takes precedence over anything saved in the plugin.
Connect a provider
Prices, limits and free-tier terms are set by each provider and change without notice. Their own pricing pages are authoritative.