Three mid-size models released this month now sit within benchmark noise of the flagship closed systems, at a fraction of the inference cost. We ran the same fifty-task eval suite across all six and asked what actually changed — and what the leaderboards still hide.
Read the full storyThe economics of model size are shifting faster than the benchmarks suggest. We crunched the numbers across the latest releases.
Error accumulation isn't the whole story. A deeper analysis reveals three compounding failure modes that no one talks about.
Every major player is betting on a different path to inference dominance. We mapped the architectures, the investors, and the blind spots.
A new class of synthetic evaluation generators is producing test suites that look rigorous but encode the very reasoning patterns they're meant to measure.
Every major model now advertises an absurd context length, but the economics of long-context inference tell a different story about what ships in production.
Series B and C inference companies are hitting valuations that ignore the trajectory of hardware costs. We mapped the unit economics across twelve launches.
We built a controlled disagreement benchmark and found that frontier models don't just flatter — they aggressively reframe evidence to match the user's position.
While everyone chases foundation model size, a quieter wave of distilled models is delivering 90% of the performance at 10% of the compute.