Perplexity has added a new feature to its Mac app called Hybrid Compute. It splits a single task in two: the sensitive parts run on your own computer using a smaller AI model, while everything else goes to a powerful cloud model like Opus 5 or GPT 5.6. The goal is to keep private information off someone else's server while still getting the benefit of a top AI model for the rest of the work.
The example Perplexity gives is a lawyer preparing a brief. The tool can compare a case against public case law using the cloud model, while keeping the client's confidential details on the lawyer's own Mac. Perplexity built a privacy classifier that automatically flags files or text it thinks should stay local, and the user reviews that decision before anything gets sent anywhere.
This matters more than it sounds. Courts have already started treating AI use itself as a legal risk: a federal court ruling in United States v. Heppner found that using a consumer AI tool destroyed attorney-client privilege, not because of a leak, but simply because the tool's terms of service allowed the company to retain and train on the data. For any business handling contracts, health records, or financial statements, that is not a hypothetical problem.
The cost angle is real too, but secondary. Cloud AI models charge based on usage, and anything handled by the local model on your own Mac costs nothing since there is no server bill involved. For companies running lots of tasks through Perplexity's $200-a-month Computer product, offloading routine work to a free local model adds up quickly.
Perplexity is honest about the tradeoff, which is worth respecting. Its own team said a fully cloud-based result will almost always beat a hybrid one, because the smaller local models are simply less capable. Hybrid Compute is a compromise for people who care more about privacy or cost than about getting the single best output every time.
This idea did not start with Perplexity. Apple has run a similar setup since 2024 with Private Cloud Compute, keeping simple requests on the iPhone and sending only complex ones to Apple's own servers, and Microsoft's newer laptops and Google's Pixel phones do versions of the same thing. What Perplexity has done is bring that local-versus-cloud split into a general-purpose work agent and hand the user direct control over where that line sits.
For now, this only runs on Apple Silicon Macs with macOS 15 and a decent amount of memory, and it is limited to paying subscribers and enterprise customers. The bigger signal is where AI tools are heading: businesses handling sensitive data will increasingly need to decide, task by task, how much privacy and cost they are willing to trade for a better result. That decision used to be simple, just use the best AI available, and it no longer is.