Does the quality of the model used result in siginificant difference for flutter-skill? What are the suggested models?

[finnvyrn]

As in the title.

[charliewwdev]

Yes, model quality makes a significant difference with flutter-skill. Here's what we've observed:

Recommended models (ranked):

1. Claude Sonnet 4.5 / 4.6 — Best overall. Excellent at reasoning through UI hierarchies, writing multi-step test flows, and recovering from errors without hand-holding. Handles the full 130-tool context well after connecting.

2. GPT-4o — Good, but tends to over-call screenshot and inspect when it should just act. Works well for simple automation flows.

3. Gemini 1.5 Pro / 2.0 Flash — Works reliably after v0.9.29+ (which fixed the 128-tool limit for Vertex AI). Flash is surprisingly capable for cost-sensitive tasks.

4. Claude Haiku 4.5 — Best price/performance for repetitive tasks (form filling, regression flows).

Key insight: The bottleneck is not raw intelligence — it is whether the model understands when to stop inspecting and just tap. Claude models handle this best in practice.

Tip (v0.9.34+): flutter-skill now sends notifications/tools/list_changed after connecting, so agents that support MCP notifications (Cursor, Claude Code, Cline) automatically expand from 13 connection tools to all 130 automation tools.