llm-gemini adds Gemini 3.8 Flash with thinking levels

The async fix is the actual production concern here, not the model support. Willison's plugin now properly tracks resolved versions in concurrent workflows, addressing a reliability gap that matters far more to enterprise adoption than which model flavor ships first.

This sits directly alongside the enterprise consolidation pattern documented in the self-hosted LLM piece from yesterday. That story showed how one company solved fragmentation by building a single internal model covering 200+ use cases. Willison's approach is the opposite strategy: standardize the interface layer (the CLI) so developers can swap between Google, Anthropic, and OpenAI models without rewriting code. Both solve the same underlying problem (model sprawl), but one consolidates downstream while the other abstracts upstream. The tension matters: if abstraction layers become reliable enough, enterprises may skip the expensive post-training work and just route requests to the cheapest frontier model that meets latency requirements.

If llm-gemini's async tracking becomes the de facto standard that other CLI wrappers (Claude, GPT) adopt in the next two months, that signals the abstraction layer is winning. If instead each model provider ships their own CLI with proprietary async handling, the fragmentation persists and enterprises default back to consolidation strategies like the self-hosted approach.