Keeping your models current

AI companies release new models often. You don't have to wait for a new version of Cerevisor to use them: press Refresh models and Cerevisor asks your provider which models your account can use today. New ones appear straight away in every model picker: on agent cards, in the chat builder and in the provider settings. You don't need to restart.

Where it can find them, Cerevisor also fills in each model's price, how much text it can read at once, its reasoning levels and whether it accepts pictures.

Refreshing

Open Your model providers (Settings → Providers, or the provider button on Home).

• Refresh models, under each saved provider, checks that one provider.
• Refresh all models, at the top, checks every provider you have switched on, two at a time.

You can also refresh from an agent card: open its model picker and press Refresh models beside Use default. It checks the provider shown in the picker, and new models appear in the open list. The picker stays open and your choice doesn't change.

A short line under the provider tells you what happened:

• "2 new models: gpt-6.1-sol, gpt-6.1-luna.": the new models are ready to pick.
• "No new models. Checked 14:02.": nothing changed since the last check.
• "Couldn't refresh: … Showing the list from 28 Sept.": the provider couldn't be reached (you may be offline, or need to sign in again). The last good list stays in place and you can keep working with it.

A failed or empty answer never replaces a good list. Nothing opens a dialog, and your pickers keep working while a check runs. When a picker shows the last good list because the provider didn't answer just now, it says "Showing the list from 28 Sept."

Editing a provider, for example switching it off and on or choosing its default model, keeps its list. Only a new key, a new address or a new sign-in method makes Cerevisor look the models up again from scratch.

What a refresh never changes

A refresh only changes which models you can choose. It never changes a model you've already chosen: not on an agent, a workflow, a provider's default, a custom model name, or your own prices. If a model you use disappears from a provider's list, it stays selected and visible, and the run checks it as usual.

Updating Codex, Claude Code and Grok

Three providers get their model list from a program on your computer. That list can only be as new as the program. When a newer version exists, Refresh tells you, for example:

"Codex 0.159.2 is available; you have 0.155.0. New models may need it."

Press the Update button beside it. Cerevisor shows the output as it goes, then checks the models again.

• Codex. Cerevisor updates the Codex it actually uses, the same way that Codex was installed. A Codex installed with npm is updated with npm. Any other Codex is updated by Codex's own updater, so you don't need a terminal. Cerevisor never installs a second copy. Only a very old Codex can't update itself. For one of those, Cerevisor tells you where it is, so you can update it the way you installed it and then press Refresh models.
• Claude Code. Update Claude Code downloads Claude Code into Cerevisor's own data folder. It never changes the copy that came with the app. Before using the download, Cerevisor checks it against the fingerprint published with it, then makes sure it starts and lists its models. If any check fails, the download is deleted and the Claude Code you had keeps working. Cerevisor then uses the newest Claude Code it can find: your own, its downloaded copy, or the one that came with the app.
• Grok. Update Grok runs Grok's own updater for you, so you don't need a terminal.

Nothing is installed or updated until you press the button.

What Cerevisor contacts, and when

Cerevisor contacts these places only when you press Refresh models, Refresh all models or an Update button. Opening a model picker contacts none of them.

• Your providers, to ask which models your account can use.
• OpenRouter's public price list, for prices and other details of models from OpenAI, Anthropic, Google and xAI. This is a public page: no key and nothing about you is sent.
• npm, the public registry these tools are published on, to see whether Codex, Claude Code or Grok has a newer version. Update Claude Code also downloads from it.
• Artificial Analysis, only if you've added your own key (below).

None of these requests carries your workflows, your files, your keys (except the Artificial Analysis key, sent only to Artificial Analysis) or anything that identifies your computer.

Models running on your own computer or local network, such as Ollama, LM Studio or a server at a local address, are never looked up anywhere. Their details stay exactly as your server reports them.

The red warning: "Your spending limit can't see this model's cost"

Spending limits (an agent's limit, a workflow's limit, a Factory mode item's limit, or Operator's daily limit) can only count what they can price. When Cerevisor has no price for a model, what that model spends is not counted against your limit. When that happens, Cerevisor warns you loudly:

• before the run, at the top of the summary you confirm;
• during the run, as a red card in the run window when an agent starts on that model, or when smart routing or a fallback picks it;
• in the cost line, which reads, for example, "$1.20 + usage on gpt-6.2 with no known price".

The run still goes ahead. The warning only tells you that the limit can't protect you from that model's cost.

What to do:

1. Press Refresh models for that provider. Prices often appear on OpenRouter's list a little after a model's release.
2. If there is still no price, you can enter one yourself in that provider's settings, if it offers custom prices. A price you enter always wins.
3. Or keep an eye on the run's cost, pick another model, or contact support@cerevisor.com.

The warning only appears for models billed per use with no known price. Models on your own computer are free. Plans you pay for monthly (a ChatGPT or Claude subscription) aren't billed per use. For Anthropic's models, Cerevisor counts an unknown model at Anthropic's most expensive rate, so your limit still applies.

Optional: benchmark scores

Smart routing picks a model for each agent set to Auto. For a model it doesn't know yet, it estimates how capable the model is from its price. If you add a free Artificial Analysis key, it uses their benchmark scores instead, which are usually a better guide.

1. Get a free key from Artificial Analysis (Get a free key under Benchmark scores in Your model providers).
2. Paste it into Artificial Analysis key and press Connect. Cerevisor checks the key once before saving it.
3. Press Refresh all models. The scores are read when you refresh.

The key is kept encrypted on your computer and sent only to Artificial Analysis. Disconnect deletes it. In the run window, a model placed this way says "tier from benchmark scores (Artificial Analysis)". A model placed by price says "tier estimated from price". A model smart routing couldn't place at all is named under "Not used by smart routing".

When you still need a Cerevisor update

Refreshing brings in new models, prices and details. A Cerevisor update is still needed when a model requires a completely new way of being called, such as a new setting or a new kind of connection.