The acceleration question has moved from speculation to measurement. Epoch AI’s new report, “Have AI Capabilities Accelerated?”, examined four core metrics and found three of them showing non-linear growth patterns, a technical way of saying capabilities aren’t just increasing, they’re increasing faster over time.
What the Report Found
Epoch AI tracked capability growth using two primary measures: its own Epoch Capability Index (ECI), which aggregates performance across a standardized benchmark suite, and the METR Time Horizon metric, which measures how long an AI system can autonomously operate on complex tasks without human intervention. Both the ECI and the METR Time Horizon showed non-linear growth, the pattern of acceleration the report identifies. The report assessed four metrics in total; three demonstrated this pattern.
Epoch AI’s findings are attributed to Epoch AI. They represent the organization’s analysis of capability trends based on its own measurement methodology. They’re not the same as a scientific consensus, they’re rigorous independent research from the field’s most credible measurement organization. That distinction matters when communicating these findings: this is Epoch’s conclusion, based on Epoch’s methodology, with high credibility given Epoch’s track record.
Context: Building on Prior Coverage
Epoch AI’s ECI framework was covered in depth in a prior TJS brief, “The Third-Party Scorecard Arrives: What Epoch AI’s ECI Means for Frontier Model Selection.” This report builds on that framework by applying it to a longitudinal question rather than a single-point evaluation. Readers familiar with the ECI framework have the context to interpret this finding; readers who aren’t should start with that prior brief.
One data point from this report connects to a prior TJS brief: Epoch AI’s evaluation placing Meta’s Muse Spark at ECI 154, between GPT-5.2 and Opus 4.6, was previously published in “Meta MuseSpark Earns Independent Epoch AI Validation.” That score appears in this report as contextual data. It’s not new ground here; the acceleration finding is.
A technical paper with arXiv ID 2604.15306 is referenced in context with generalization studies related to this research area. The paper’s relationship to the Epoch report is contextual; it’s not the primary source for the acceleration findings. The Epoch report itself is the primary source.
Why the Non-Linear Finding Matters
Linear growth is manageable. Planning, resource allocation, and governance frameworks can adapt to predictable rates of change. Non-linear growth, where the rate itself is increasing, is harder to plan around. Regulatory timelines, safety research roadmaps, and infrastructure investment cycles all assume some predictability about where AI capabilities will be at a future date. If the acceleration finding holds, those assumptions need revisiting.
This isn’t a cause for alarm framing. It’s a data point for calibration. Epoch AI’s methodology is the most credible available for this type of trend analysis. The finding should be taken seriously by anyone making multi-year decisions that depend on AI capability assumptions.
What to Watch
Whether other independent research organizations, using different methodologies, replicate the non-linear growth finding. Whether the fourth metric (which did not show acceleration in this report) provides a meaningful counter-signal or represents a different capability dimension. And how frontier labs respond to the report publicly, given that their development timelines and safety research commitments are implicitly stress-tested by an acceleration finding.
TJS Synthesis
Epoch AI has done the measurement work that most organizations are still debating whether to commission. Three of four capability metrics showing non-linear growth is a finding that deserves to sit in the planning assumptions of every organization making AI infrastructure, governance, or competitive decisions over the next two to three years. The question isn’t whether to believe it, Epoch’s methodology earns serious weight. The question is what it means for the timelines on which those decisions are being made.