Google announced Gemini 3.7 Flash on August 13. Simon Willison’s llm-gemini 0.33 shipped support for it on August 13. That is not a coincidence, and it is not a story about one developer being fast. It is what model releases look like now: the announcement and the ecosystem patch arrive together, because the announcement is mostly a version bump with a launch post attached.
Look at what else 0.33 picked up in the same release. Not just 3.7 Flash, but gemini-3.6-flash and gemini-3.5-flash-lite, plus two embedding models, gemini-embedding-2 and gemini-embedding-001. Willison notes it had been a while since the last release. So the backlog of unsupported Google models accumulated to five before anyone had reason to cut a new plugin version. Three generations of Flash, all still live, all still callable, and until this week one plugin’s users could reach none of them.
Version numbers that no longer mean anything
There is no useful mental model for the gap between 3.5 Flash Lite, 3.6 Flash, and 3.7 Flash. Point releases used to signal a modest tune-up on a known base. Now they arrive close enough together that the people building on top of them cannot run a proper eval between drops. You pick a model, you ship, and six weeks later there is a newer digit sitting next to it in the dropdown with no public account of what changed for your workload.
The honest consequence is that most teams will not switch on evidence. They will switch because the number is bigger, or they will not switch at all and quietly run 3.5 Flash Lite for two years because it was cheap and it worked. Both of those are rational. Neither is evaluation.
The plumbing is doing the heavy lifting
The more interesting line in the 0.33 notes is the LLM 0.32 compatibility work. That unlocks reasoning traces in the plugin, which means you can actually see what the model did before it answered. That is worth more to a working developer than a tenth of a point on the model name. Reasoning visibility is how you debug a bad output instead of guessing at your prompt.
It says something that this arrives through a third-party plugin in a Python CLI ecosystem rather than as the headline feature of the vendor’s own launch. Google ships the weights and the API. The layer where you compare, inspect, and swap models is maintained by people who are not Google, on their own schedule, for free.
So the practical question for anyone on Gemini right now: do you have any way to tell whether 3.7 Flash is better than 3.6 Flash for the thing you actually do? The plugin will let you call both. It cannot tell you which one to keep.