During the twelve months beginning 2029-01-01, by how many points will the highest Epoch Capabilities Index score held by a model from OpenAI, Anthropic or Google DeepMind rise?

rationale:

The median estimate of 16.0 ECI points anticipates a capability rise tracking slightly above the current historical trend, buoyed by the accelerating impact of AI-assisted research. The distribution is markedly right-skewed, with the middle 50% spanning 12.5 to 21.5 points and the p90 at 28.5 points, reflecting the potential for recursive self-improvement to trigger non-linear capability gains by 2029. Evaluating this timeline alongside other capability scaling metrics, anticipated shifts in enterprise AI pricing, and near-term lab risk assessments leaves the central expectation near 16 points while supporting slightly wider extremes. Conversely, the p10 of 9.5 points captures the risk of physical compute bottlenecks, funding plateaus, and safety-driven pauses.

Recent frontier movements anchor the base rate. By October 2026, Claude Opus 5.5 and GPT-6 Astra hold the top ECI scores at roughly 167 and 166.5, respectively https://epoch.ai/benchmarks. Moving from GPT-5 (anchored at 150 in August 2025) to 167 over 14 months translates to an annualized gain of about 14.6 points. Epoch's own breakpoint analysis has placed the frontier capability growth at 15.5 points per year since April 2024 (90% CI 13–18) AI capabilities progress has sped up AI capabilities progress has sped up, and its September 2026 update estimated a steady linear gain of roughly 14 points per year since the introduction of reasoning models https://epoch.ai/data-insights/eci-frontier-trend https://epoch.ai/data-insights/eci-frontier-trend. A baseline forecast naturally centers heavily on this established 14–16 point band.

The primary driver of upside variance is the automation of AI research. Anthropic reports that Claude already leads 26% of its model R&D work under human supervision As AI companies get closer to 'recursive self-improvement ... and writes over 80% of merged code When AI builds itself. Meanwhile, OpenAI has publicly targeted March 2028 for the deployment of an automated AI researcher Inside the Race to Make AI Build Itself. If these milestones yield compounding efficiencies, calendar 2029 could coincide with the onset of partial recursive self-improvement. While lab testimony carries incentive biases, the timeline alignment makes a step-change in capability growth highly plausible, justifying a fat right tail that pushes well past 25 ECI points.

Significant headwinds shape the lower bounds of the distribution. Compute scaling may slow as power constraints and longer build lead times stretch development cycles Compute scaling will slow down due to increasing lead times How much power will frontier AI training demand in 2030?. Analysts also point to a potential funding plateau around 2028–29, which could compress compute growth from its current ~3.55x annual pace down to roughly 1.4x Slowdown After 2028: Compute, RLVR Uncertainty, MoE .... In addition to physical constraints, labs might deliberately throttle capability releases due to safety concerns.

Finally, structural features of the Epoch Capabilities Index introduce unavoidable noise. Because Epoch continuously refits the index as models are evaluated, historical snapshot scores shift retrospectively—for instance, GPT-6 Astra's score dropped 2.9 points within ten days of launch as more software-engineering benchmarks were processed https://epoch.ai/data-insights/astra-eci-breakdown Epoch Capabilities Index (ECI). Furthermore, benchmark saturation can temporarily compress observed gains before new evaluations are integrated Epoch Capabilities Index (ECI). While the forecast targets the true underlying rise in capabilities, the discrete timing of frontier model releases near the January 1 or December 31, 2029 boundaries can easily shift the measured calendar-year rise by ±3–4 points, predictably widening both tails of the distribution.