For the past two years, practitioners have felt that AI is moving faster. Epoch AI now has data that quantifies the feeling.
According to Epoch AI’s updated analysis, the rate of frontier model improvement nearly doubled following an inflection point in early 2024. Before the break: approximately 8.2 points per year on Epoch AI’s Epoch Capability Index (ECI). After: approximately 15.5 points per year. The pattern is piecewise linear, meaning the acceleration isn’t spread smoothly across years, it snapped upward at a specific point.
Epoch AI’s analysis points to reinforcement learning on reasoning tasks and increased compute investment as contributing factors. The Wire flags this attribution as a source inference rather than a direct Epoch AI conclusion, so treat it as directional rather than definitive. The ECI growth rate figures, 8.2 and 15.5, are Epoch AI’s reported metrics from their proprietary tracking methodology. They reflect Epoch AI’s composite measurement across frontier models, not any single benchmark score.
That distinction is worth pausing on. ECI is not HLE. It’s not SWE-Bench. It’s Epoch AI’s own composite index for tracking capability across the frontier. If you’re reading this alongside the Claude Opus 4.7 brief from today’s package, the two pieces of data are related but not equivalent: the ECI tracks aggregate frontier progress, while model-specific benchmarks measure individual releases against defined tasks.
What does a near-doubling of capability growth rate mean practically?
For developers and technical teams: the half-life of a “current best” model is shrinking. A capability threshold your team evaluated six months ago may no longer represent the frontier. Build your evaluation cadence accordingly.
For compliance teams: this is the data point regulators may not yet have fully integrated into threshold-setting. The EU AI Act’s compute-based thresholds were calibrated against a capability growth rate that Epoch AI’s data suggests has meaningfully changed. If capability per unit compute is increasing, which is what accelerated ECI scores imply, then FLOP-based classification boundaries may need recalibration. This is worth watching for EU AI Act compliance planning. See the Regulation pillar for ongoing EU AI Act coverage.
For infrastructure and investment observers: faster capability growth drives demand. More capable models require more training compute, more inference infrastructure, and more evaluation tooling. The acceleration Epoch AI describes is consistent with the infrastructure investment thesis behind recent AI chip funding. See the Markets pillar for AI chip and infrastructure funding context.
One note on methodology transparency: Epoch AI’s ECI is a proprietary metric. Its construction reflects methodological choices that aren’t publicly documented at the same level as established academic benchmarks. The figures here are useful as directional signals from a credible research organization, not as independently reproducible measurements. Epoch AI publishes its research openly, and their track record on capability forecasting is respected in the field. But readers should distinguish between ECI as an analytical tool and peer-reviewed benchmark results when making decisions that hinge on precision.
What to watch: whether the inflection holds through Q2 2026 releases, and whether regulatory bodies begin referencing external capability indices in threshold-setting discussions.