Hey, thanks for the question. From your screenshot, I can see the Task subagent Draft architecture design spec ran on Sonnet 4.6 Medium even though you didn’t pick it. Subagent models are configurable, here’s how:
Custom subagents .cursor/agents/*.md, like your Design Spec. Add model: in the YAML frontmatter:
---
name: Design Spec
model: claude-sonnet-5-thinking # or inherit to use the parent chat model
---
Bracket syntax pins the variant and effort, for example model: claude-opus-4-8[effort=high]. Note: this syntax isn’t officially documented yet, and it doesn’t work with the CLI --model flag. For CLI you need to pass the exact slug, for example claude-opus-4-8-thinking-high.
Built-in subagents, Explore, etc. Go to Settings > Agents > Subagents and pick a model per subagent type.
At the prompt level. You can tell the parent agent use my current model for any subagents. The Task tool also accepts a model argument.
About Grok automatically swapping in a Sonnet slug for Task calls without being asked, that’s known behavior. I can’t share an ETA for a change yet. If the model still doesn’t apply after setting model: in the frontmatter, send the Request ID and I can take a closer look.
Pricing. Sonnet 4.6 and Sonnet 5 are both $3/M in and $15/M out at standard pricing, but Sonnet 5 currently has a launch promo of $2/$10 through August 31, 2026, so it’s cheaper right now. Keep in mind actual cost depends on how many tokens a given task uses, not just the per-token rate.
More details in the related thread: Subagent Model Choice Not Respected