Every week has a new model. This week had three worth noticing, and none of them were about being the biggest.
Local agents got real
Meta released Muse Glimmer, a 30-billion-parameter model built for always-on local agent workflows, with open weights under Apache 2.0. Meta says it's small enough to run on a Mac or PC with a single consumer GPU, and it's aimed at local agents, function calling and coding [1].
Why I care: a capable model that runs on hardware you own means data that never leaves the building. For a lot of organizations with privacy or compliance worries, "local and good enough" beats "cloud and amazing."
The workhorse got cheaper
Google introduced Gemini 3.7 Flash just three weeks after 3.6 Flash, with an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, half the original 3.6 Flash cost. That runs through December 31, 2026, before rising to $1.50 and $7.50 [2].
Translation: the cost of having AI do routine work keeps falling. Budget conversations that didn't pencil out in the spring might now. (Also: set a calendar reminder for January 1. Introductory pricing is called that for a reason.)
Microsoft brought its own
Microsoft AI previewed MAI-Thinking-1, a medium-sized reasoning model available in Microsoft Foundry. Microsoft says it's among the strongest in its weight class and emphasizes "clean, traceable" training data without distilling from other labs [3]. Whatever you think of the benchmarks, more competition among the big platforms is good news for buyers.
The practical read
- Stop picking "the best model." Pick the cheapest model that does the job reliably, and test that claim on your own work.
- Revisit local AI. If privacy has kept you on the sidelines, the hardware bar just dropped.
- Watch the pricing fine print. Intro rates, regional availability and preview status all matter for anything you build on.
The model race used to be about who's smartest. It's quietly becoming about who's cheapest, smallest and easiest to trust. That's a race the rest of us actually benefit from.
How I'd test one
Pick twenty real tasks your team does every week, write down what a good answer looks like, and run each candidate model against them. Score them yourself. It takes an afternoon, and it will tell you more than a month of reading launch posts. Repeat when prices change, which, judging by this week, is roughly every three weeks.
Keep the twenty tasks in a shared doc. They become your team's own benchmark, and they're worth more than any leaderboard.