GPT-6 lands and nobody can verify anything
OpenAI apparently launched GPT-6 (also called 'Astra') today, with early reporting claiming 98.6% on ARC-AGI-3, compared to roughly 30% from previous frontier models. The HN thread was a mess of dead links and skepticism, with commenters unable to confirm the benchmark claims or even access the announcement.
This is the clearest example yet of model release velocity outpacing the community's ability to evaluate what's actually being shipped. Benchmarks like ARC-AGI-3 are themselves new enough that most people don't have intuitions about what 98.6% actually means in practice. The gap between a press release number and a usable model assessment is widening.
A separate thread noted 'model fatigue' directly: one commenter said new models are coming out '10x faster than new JavaScript frameworks were coming out 10 years ago.' The difference, they noted, is that models are easier to adopt than frameworks. But easier to adopt is not the same as easier to understand.
So what?
If you're building on top of any specific model, the ground is shifting under you faster than you can ship. Founders should be designing for model-agnosticism now, not after the next surprise release forces the issue. The benchmark arms race also means marketing numbers are increasingly useless for product decisions.