The distance between those two numbers is the most honest thing published about AI progress this year.

When six labs are within 5% of each other, the question stops being which model is best and becomes which one you can afford to run.

SWE-bench Verified went from 60% to near 100% in a year. A test everyone passes measures nothing.

Compute spent thinking can substitute for compute spent training. That trade reorganised the field, and moved the bill onto you.

Agents that work for five steps fall apart over fifty. The research says the problem is what they are carrying, not how clever they are.

Publishing weights used to mean you had lost the frontier. Now it means you want the distribution.

Daybreak now runs Blue and Red tiers behind approval. The gate says more about the capability than the launch does.

Google says its new entry-level model beat Anthropic and OpenAI across nine benchmarks. The cadence matters more than the scores.