The AI models screen looks like the place where you pick a model, and most of the time that reading is close enough. It lists a handful of models with capability, speed and cost tiers, a sentence on what each is good for, and a sentence on the trade-off you are accepting. Somewhere near the top of a row there may be an active badge.
Then one day there is no badge anywhere on the page, the AI assistant is plainly working, and the screen appears to be telling you that you are running nothing at all.
That is worth understanding properly, because the fix is almost never the one people reach for. The list on that page is a fixed reference card written into the product. It is the same five rows for every account on every plan, and it does not change when your account changes. It is not a query against your provider account, and it is not a list of what you are entitled to call. It is a menu printed on the wall. The badge is the only part of it that knows anything about you.
What the active badge is actually testing
The badge is decided per row by two narrow rules, one for each provider, and both are stricter than they look.
An Anthropic row is marked active when two things are true at once: your account has an Anthropic key saved, and that row's model identifier is exactly the model your account is configured to use. If you have never typed a model name, the configured value is treated as Claude Sonnet, so the Sonnet row lights up on a fresh account with a key in place.
An OpenAI row is marked active only when your account has no Anthropic key, does have an OpenAI key, and the row is one of the two marked as a default. Every other row on the page is inert by construction: it can describe itself, but it can never claim to be the one running.
One key silently outranks the other
Read those two rules together and a precedence falls out that the screen never states in words. The moment an Anthropic key is present, the OpenAI rows can no longer be active, whether or not an OpenAI key is also saved. The OpenAI side of the page only comes alive on an account that has an OpenAI key and no Anthropic key.
This matters most in the situation where people are most likely to misread it: you saved an Anthropic key months ago, you have since added an OpenAI key for something else, and you are now looking at the page trying to work out which one the assistant is really using. The page is not ambiguous, but it answers by omission. The OpenAI rows going quiet is the answer.
If you want the OpenAI side to be the one in use, the change is to remove the Anthropic key rather than to add anything. Which key is present is the switch, and there is no separate preference to set.
Type a model that is not on the card and nothing lights up
Here is the case that produces the blank page, and it is the common one on accounts that have been tuned by somebody who knew what they were doing.
The configured model is free text. You can put any model identifier your provider accepts into it. The badge, though, is decided by an exact match against the three Anthropic rows printed on the card. Put in a model that is newer than the card, or a variant the card does not list, and no row matches, so no row is marked active.
Nothing is broken at that point. Your assistant goes on calling the model you named, because the setting is what the assistant reads and the card is only what the dashboard prints. The screen has quietly stopped being able to describe your configuration, which is a display limitation rather than an outage. The tell is that the page is blank of badges while the usage table below it is busy.
It is worth knowing this before you go and change a working configuration to something the page recognises. A blank badge is not evidence of a problem, and treating it as one is how a tuned setup gets reverted for no reason.
The table underneath is the half that cannot be wrong
Below the catalog is a usage table, and it is built the opposite way round. Nothing about it is written in advance. It groups the AI calls your account actually made over the last thirty days by provider and model, keeps the ten busiest, and reports for each one the number of calls, tokens in and out, how many ended in an error, the mean latency, and when it was last used.
That is a record of what happened rather than a description of what is available, which makes it the authoritative answer to the question people bring to this page. If you want to know which model you are running, do not read the badges. Read the top row of the usage table.
The table has its own honest silences. It is capped at ten rows, so a long tail of occasional models is not shown. Anything that never recorded a provider or model name is grouped under unknown rather than dropped. And if the underlying usage table has not been deployed on your installation, or the AI has simply not run in thirty days, the query comes back empty rather than failing, so you get an empty table and no error message.
That last point makes an empty table ambiguous in exactly the way a blank badge is not. Empty means either nothing ran or nothing is being recorded, and the two look identical from here.
Reading the two halves together
Once you know which half is which, the page becomes a decent diagnostic, and the combinations are readable.
A badge and matching traffic in the table is the healthy case: the screen understands your configuration and the configuration is being used. No badge but real traffic means you are running a model the card does not list, which is fine and usually deliberate. A badge but no traffic means the configuration is understood and nothing has called it, which points at the assistant being switched off, or at billing rather than at model choice.
Nothing anywhere is the case worth pausing on, because it is genuinely ambiguous. Take it to whether AI is enabled and whether a key is saved, and only then come back to model choice.
The tier labels deserve one caution. Capability, speed and cost are editorial judgements written when the card was written. They are useful for a first pass and they are not live pricing. Treat them as a starting opinion, and let your own usage table and bill settle any argument about cost, because those reflect your traffic and the card cannot.
How MyLiveChat fits
The models screen is a reference card with one live badge, sitting above a usage table that reports what your account really did. Read it in that order and it answers the question quickly. Read the badge as an inventory and it will mislead you on any account that has been tuned.
Model choice is one of the few genuine cost levers in the AI assistant, which is why the page pairs the two halves at all. What actually drives your AI chat costs covers the levers that move the bill more than model choice does, and managed AI replies or your own API key explains the key arrangement that this page's precedence rule is quietly reading.
If you are weighing a switch, decide it on evidence rather than on tier labels. Testing AI chat answers with an eval suite is how to compare two models on your own questions before you commit, which is the only comparison that reflects your customers.