Charts

Price of a fixed capability over time

A GPT-4-class answer costs roughly 100× less than it did at launch — and frontier-class is following the same slope.

List price vs. effective price per model

Cached and batch discounts routinely cut the sticker price in half; the gap is widest on the newest frontier models.

Days to open-weight parity

Open-weight models now match each frontier capability jump in months, not years — and the gap keeps shrinking.

Days from launch to first price cut

Labs cut prices faster than they used to. Mid-tier models get discounted within a quarter; flagships hold longer.

Who holds the paper on which data center

Frontier labs sit at both ends of the lease: they buy capacity from clouds and lessors who, in turn, buy chips that flow back into those same labs' clusters.

Release-to-deprecation span per model

Frontier models stay current about 6–12 months before a successor or a quiet deprecation. Open-weight lines turn over faster.

Attention half-life after a launch

Social mention volume after a major launch decays by half within ~18 hours. The second day is already a different story.

Weekly developer sentiment by lab

Developer chatter swings hard around launches and outages. The index is noisy week to week — the trend across quarters is the signal.