Faster Background Tasks + Refreshed Recommended Models

April 15, 2026

The local model that does the boring work behind the scenes — generating conversation titles, extracting memories, running the web agent — switched to a faster one that's already kept warm on the GPU. So titles appear almost instantly, memory extraction is tighter, and the web agent moves faster on browser steps. The "recommended" badges in the model picker were also refreshed to reflect what's strongest right now.

What's better

• Background tasks are quicker. Title generation, memory extraction, and the web agent all run against a model that's already loaded, instead of paying a cold-start each time.
• The model picker has fresh "recommended" picks for general work, surfacing better quality-per-dollar than the previous defaults. Open it and you'll see the new badges.

When you'd notice it

• You start a new conversation and the title appears almost instantly instead of taking a few seconds.
• You're using the agentic browser. Each step (click, read, decide) feels snappier — no round-trip to an external API just for navigation decisions.
• You open the model picker to choose something for a session. The recommended list looks different — try one of the new picks on a quick task and see how it feels.

Heads up

A couple of frontier-tier models that used to carry the "recommended" badge no longer do. They're still fully available; they just aren't surfaced as the default starting point. For most general-purpose work the new recommended picks are better value. For genuinely frontier-tier prompts, the picker still has Opus 4.7, Sonnet 4.6, and GPT-5.4 — pick them deliberately.