The story everyone expected was a slowdown. The data shows the opposite. Epoch AI measures frontier capability rising about 15.5 ECI per year, with a 90 percent interval of 13 to 18, and notes the rate has grown faster since early 2024. The curve that many predicted would bend toward a plateau bent the other way.
The acceleration comes from several inputs compounding at once: training compute up about 5 times a year, algorithms 3 times more efficient annually, and a wave of investment, with frontier labs having raised more than 170 billion dollars. When several exponentials stack, the combined capability curve steepens rather than settling.
What an acceleration does to forecasts
The trap is forecasting a rising curve with a fixed yearly rate. If the rate of improvement is itself climbing, as it has since the 2024 inflection, a constant-rate forecast undershoots every single year. Plans that assumed diminishing returns have been consistently wrong on the low side, which is an uncomfortable kind of error to keep making.
How to plan against a moving rate
- Do not anchor on a plateau. The expected ceiling has not shown up in the data, a caution we raised in the scaling curve nobody wants to extrapolate.
- Shorten your horizon. Replan capability assumptions every two quarters, not every two years.
- Separate capability from value. Faster models do not automatically mean faster returns, as the agentic cancellations show.
Betting on a slowdown has been the losing trade for two years running. The capability growth estimates come from Epoch AI.
Why a faster rate changes product risk
An accelerating capability curve changes the risk of every long product bet. If models improve at a steady pace, a roadmap can assume that today's hard problems will become easier in a predictable sequence. If the rate itself rises, the ordering can change. Features that seemed impossible at planning time can become ordinary before the project ships, while safeguards designed around older model limits can become stale.
That matters most for teams building around absence. Some products are viable because models cannot yet perform a task cheaply, run locally, or reason across a large enough context. A faster curve shortens the life of those assumptions. A compliance tool built around document summarization, for example, may face new competition once long-context models can ingest the full source file. A research workflow built around manual synthesis may look different once models handle larger evidence sets with lower error rates.
The planning mistake is to treat capability as a background variable. It should be a line item in the roadmap. Teams need to ask which current product choices depend on model limits and how those choices will age if the next release is better than expected. The answer may change pricing, hiring, integration depth, or the decision to build a feature at all.
The data is strong, but the interpretation is narrow
Epoch's ECI measure is useful because it tries to summarize the capability of frontier systems over time. It does not mean every product sees a 15.5-point annual improvement in value. Model progress arrives unevenly. Coding, visual reasoning, tool use, long-context recall, and scientific problem solving can move at different speeds. A broad index is a climate reading, not a weather forecast for each application.
That caveat matters because business cases often translate capability into revenue too quickly. A model can be more capable and still fail a workflow if it is too slow, too expensive, too hard to audit, or too unreliable in rare cases. The acceleration raises the ceiling. It does not remove the need to test the floor.
The best interpretation is probabilistic. A faster frontier increases the odds that currently marginal applications become practical sooner than expected. It also increases the odds that an internal build will be overtaken by a commodity model before the investment pays back. Both can be true for the same company.
How to update forecasts without chasing noise
Roadmaps should separate three clocks: frontier capability, deployable capability, and organizational adoption. Frontier capability is the first demo that proves a task can be done. Deployable capability is the point where it is cheap, fast, and stable enough for normal users. Adoption is the slower process of changing workflows, training staff, and adjusting controls. Acceleration at the frontier pulls on all three clocks, but it does not make them identical.
A practical forecast can use scenarios instead of a single date. The base case assumes the recent rate continues. The upside case assumes another post-2024 speed-up. The downside case assumes infrastructure, data, or evaluation limits slow the curve. Each scenario should name the product decisions that would change if it became true.
The important habit is cadence. Updating assumptions twice a year is enough for most teams and avoids reacting to every launch thread. A capability curve that keeps bending upward rewards vigilance, not panic. The organizations that adapt well will be the ones that treat model progress as a measurable input rather than a surprise.
The simplest worksheet is a dependency list. For each product line, write down which tasks are blocked by model quality, which are blocked by cost, and which are blocked by trust. Then revisit the list after each major frontier release. That exercise turns a vague acceleration story into concrete decisions about what to build, what to buy, and what to postpone.
