Choose the model behind your chat, and what changes if you pick a global one

Until now the ScholarThread chat answered with one model that we picked for you. That is a sensible default, but it isn't the right choice for every question. A quick "what does this paper say about sample size?" doesn't need the same model as a long, careful comparison across twenty papers.

So the chat now has a model picker: you choose which model answers, for one chat or as your default everywhere. It is rolling out gradually. If you don't see a model button above the message box yet, it isn't switched on for your account.

Picking a model

Above the message box there is a small button showing the model the next message will use. Click it to see the list. Each model shows an indicative cost relative to the default model, so you can see at a glance which ones are cheaper and which ones use your AI credits faster. The estimate is rough, because the real cost of an answer depends on how much of your library it reads and how long the reply is. The pricing page has a table with the cost of every model in AI credits.

Two shortcuts sit at the bottom of the list:

• Make this my default keeps the current model for every new chat.
• Use my default drops a per-chat choice and goes back to your default.

You can also set your default in Settings → AI usage, in the Chat models card.

If a model can't be used for a chat, for example because it's no longer offered, ScholarThread tells you before your next message and switches that chat to your default rather than failing when you hit send.

Where your data goes

Most models in the list run in AWS's European Union regions, as the rest of ScholarThread's AI features do. These are the Anthropic Claude models and Amazon Nova.

Three models are different. OpenAI's GPT-6 Luna, GPT-6.1 Sol and GPT-6 Astra are marked Global in the list. They run on Amazon Bedrock with global cross-Region inference, which means AWS may process your request in any of its supported regions, including ones outside the EU, for example the United States.

That is a real difference for anyone who works with sensitive material or whose institution has rules about where data may go, so these three are optional and locked by default. Nothing is ever sent to them unless you choose one.

The consent step

The first time you pick a global model, ScholarThread shows a consent form before anything is sent. It says what the model receives when you send a message in a chat that uses it:

• your message and the earlier messages in that chat,
• facts the assistant has remembered from your chats (chat memory, including pinned items),
• passages retrieved from your library to answer the question,
• and any files you attach to the message.

It also says which models and which provider are involved, and that AWS may process the request outside the EU. You agree once, per user, and the model unlocks.

You stay in control afterwards. In Settings → AI usage you can withdraw your consent at any time. A chat that was using a global model then goes back to ScholarThread's standard model, and you'll see a short notice on that chat so it never changes silently. If you later pick a global model again, you'll be asked again.

Not giving consent costs you nothing else. Every other feature, and every EU model, works exactly the same.

What doesn't change

The global models are available only in the chat. The other assisted features, such as drafting claims, finding supporting passages and populating notes, keep running in the EU as before. As before, we don't use your content to train models, and AWS doesn't use content submitted to Bedrock to train models or share it with model vendors. Chat answers are still identified as AI-generated, the model that wrote each one is recorded, and the academic-integrity guardrails apply whichever model you choose.

The details are in the updated privacy policy, in the AI-assisted features section. If you have questions about a particular institutional requirement, write to us via support.