AI Security Institute says agent benchmarks must report performance curves over compute rather than static scores

Important framing from @AISecurityInst "An AI agent's performance is best understood as a capability curve over compute, not a single score...The takeaway is about good measurement. Evaluations should report capability curves, not single numbers."

So getting 5x performance for every 10x of inputs is much closer to linear than we usually see. This is why the time horizons are one of the only areas actually showing exponential improvements in AI performance over time (though the inputs are also exponentially increasing).

Because it is a log-log chart and those slopes are all roughly linear, the relationship is roughly a power law. The 80% time horizon is rising as the 2/3 power of the number of tokens. This means every 10x of the tokens buys about 5x the time horizon.

Since the time and money a human requires to complete a task is roughly linear with the time horizon (an hourly wage), the models here are scaling less well than humans.

But it is also worth noting that this scaling is very good by usual standards in AI scaling laws. Pre-training scaling requires 1,000,000x as much compute to halve the error. Inference scaling on maths benchmarks gives logarithmic improvement, and so does RL training.

Moreover, there isn't much improvement in how they scale over time (all the lines for different models have similar slopes). So for any time horizon, next year's models will be better, but we'd expect them to fall farther behind humans as the task extends.

This suggests there is a more efficient approach out there — a cognitive algorithm that scales as the human one does.

@tobyordoxford I agree there isn't a dramatic change in the slope steepness over time, but if you squint a bit I think a few of the earlier models (GPT 5, sonnet 4.5, Opus 4.5) do look like they have slightly shallower slopes?

@tobyordoxford The 50 day doubling time of the 50M token time horizon is pretty remarkable

@tobyordoxford @AISecurityInst Sublinear scaling. 10x+ tokens for 10x time horizons means inference compute becomes the binding constraint for production agentic systems.