Model releases are coming too fast to benchmark meaningfully

Three new model releases hit the front page today: Claude Sonnet 5, Nano Banana 2 Lite (which appears to be a small, fast multimodal model), and Leanstral 1.5 (Mistral's formal verification focused model). Each thread has commenters struggling to understand what's actually improved. On Nano Banana 2 Lite, someone notes a 'massive decrease in latency' but says the product page makes it hard to understand what changed. On Sonnet 5, commenters are comparing it to Opus 4.8 on cost versus performance. On Leanstral, someone asks whether it's useful for programs or only theorems.

The Mistral thread has a particularly sharp exchange: someone asks directly whether anyone uses Mistral because it's actually the best at something, or only 'because EU.' The honest answers are mostly the latter. That's a damning signal about differentiation in the mid-tier model market.

The pattern: the release cadence has outrun the community's ability to evaluate. Benchmarks are gamed, marketing copy is vague, and the people actually building things are making decisions based on vibes and geography more than rigorous testing.