Don't stop here
Hand-picked guides our readers explore right after this one.
Research-grade prompts for Perplexity AI's search-powered responses
Read the guideExpert guide to Claude prompts with XML tags, artifacts, and complex reasoning
Read the guideAI prompts for LinkedIn posts, profile optimization, outreach, thought leadership, and job search
Read the guideYou pick a specific model in Perplexity, ask your question, and by the next thread the selector has quietly slid back to Auto or Best. This is not a bug in your account and clearing your cache will not fix it. As of mid-2026, model selection in Perplexity is stored at the thread level in the interface state rather than at the account level, so every new thread initialises with the platform default and your previous choice simply does not travel forward. It affects Pro subscribers across browsers, devices, and accounts. There is no confirmed fix timeline, but there is one genuinely durable workaround (Spaces) plus a set of habits that stop it costing you answers.
The model selector shows Auto or Best at the start of every new thread
A model you selected yesterday is not selected today
Answers vary in depth between threads because different models actually ran
The choice persists within a thread but never across threads
Changing the setting on desktop has no effect on mobile, or the reverse
The reset happens regardless of browser, device, or signing out and in
Your selection lives in the interface state of the specific conversation. When a new thread initialises, the platform default re-engages, and there is nothing carried over from the previous thread to override it.
Perplexity does not expose a persistent default for most models. The single native exception is Sonar, which Pro users can set as a persistent default in Settings. Everything else resets.
Auto and Best are designed to pick a model per query. From the product's perspective the reset is the default working as intended, which is why it has not been treated as a defect to be fixed quickly.
Selections made in the web app do not propagate to mobile or vice versa, because the state is local to the thread in that client rather than synced as an account preference.
Because a different model actually answered, output quality genuinely differs, and users conclude their subscription has been throttled. The cause is the reset selector, not a change to the plan.
When to try: First, if you use one or two models consistently.
Open Spaces, create one per workflow (for example 'Research' and 'Quick answers'), and set the model on the Space itself. Threads started inside a Space use the Space's model, which removes the reset from your daily experience. This is the most durable available workaround.
When to try: When Sonar is an acceptable everyday model for you.
In Settings, Pro users can set Sonar as a persistent default. It is the only model with native account-level persistence, so if Sonar fits your use case this eliminates the problem outright rather than working around it.
When to try: Always, until persistence is fixed.
Make glancing at the model selector part of sending a query, the same way you would check the recipient on an email. Two seconds before the query is far cheaper than discovering afterwards that Auto answered a question you needed a specific model for.
When to try: For any multi-question piece of work.
The selection does hold within a thread. For a research session that needs a specific model throughout, keep asking follow-ups in the same conversation rather than opening new threads, and your choice stays applied.
When to try: Whenever answer quality is unexpectedly shallow.
Check the model attribution shown on the response before you rely on the output. If it is not what you selected, the reset caught you. Re-select and re-ask rather than working from an answer produced by a different model than you intended.
When to try: Before you spend an hour on local troubleshooting.
Clearing cache, switching browsers, reinstalling the app, and signing out and back in do not fix this, because it is platform behaviour rather than local state. Confirm the symptom matches the description here, then move to the workarounds instead of spending time on cleanup.
When to try: Once, and then get on with the workarounds.
Use Perplexity's in-app feedback to request a persistent account-level model default. Product priorities follow demand, and this specific request is well documented across the community but still unresolved.
When to try: When consistency between sessions is a hard requirement.
If reproducibility across sessions genuinely matters (benchmarking, consistent long-running research), run that work in a tool where the model is pinned, and keep Perplexity for search-led questions where Auto routing is acceptable.
Set up Spaces per workflow with the model pre-selected
Glance at the model selector as part of sending any query that matters
Keep related questions in one thread instead of opening new ones
Check the model attribution on the answer before relying on it
Support cannot change platform-wide behaviour, so a ticket about the reset itself will not produce a fix. Contact Perplexity if a Space that you locked to a specific model still runs a different one, if a Sonar default set in Settings does not hold, or if your Pro entitlements appear absent, since those are genuine account faults rather than the expected reset.
Because model selection is stored at the thread level in the interface, not at the account level. When a new thread initialises, the platform default (Auto or Best) re-engages and your previous choice does not carry forward. It affects Pro subscribers regardless of browser or device.
Only partly. Sonar can be set as a persistent default in Settings for Pro users, and that is the sole native exception. For any other model, the reliable approach is a Space locked to the model you want.
Yes. Threads started inside a Space use the model set on that Space, which removes the reset from day-to-day use. Creating one Space per workflow with the right model preset is the most durable workaround available.
No. This is platform behaviour rather than local state, so cache clearing, browser switching, reinstalling, and signing out and in all leave it unchanged. Use Spaces or the Sonar default instead of local troubleshooting.
Check the model attribution displayed on the response. If it does not match what you selected, the reset applied before you sent the query, and you should re-select and re-ask rather than relying on that answer.