
Introducing the Stripe Economics of AI Fellowship: The economics of AI remains surprisingly understudied. The fellowship aims to help fill that gap, by supporting grad students and early-career researchers with $, data, a conference, and community –

Assistant professor of economics @UVA; Past lives: PhD @MITEcon, @Uber, @AQRCapital, @UChicago

Introducing the Stripe Economics of AI Fellowship: The economics of AI remains surprisingly understudied. The fellowship aims to help fill that gap, by supporting grad students and early-career researchers with $, data, a conference, and community –

After Q3 earnings reports last week, we can officially update @brightertmrrw -- hyperscaler capex > cashflow
The end of the self-funded AI buildout? Hyperscaler cash capex is growing much faster than cash inflows. On current trends, they will be unable to fully fund the AI infrastructure buildout with cash from operations by the end of this year.

And if you run the numbers, in the end, a 10x increase in *total* compute gets you very roughly a ~10x increase in time horizon Basically it all comes down to this graph. The elasticity of ~1 holds across 3 OOMs

@karthiktadepall had a tweet (which I've now lost) emphasizing that the distribution of OpenAI tasks being measured shifts over time -- as the OpenAI OP writes in a footnote:

Clarification
OpenAI uses high context (at least in my reading), this is good to emphasize: first screenshot from their blog post, second from the relevant paper

My prior is anchored on exponential increases as the baseline, given the exponential increases in inputs (or empirically see metr.org/blog/2025-07-14-how…) but interesting to do this estimation
Basil, this may interest you. When I try to find best fit to the OpenAI time horizon increases, they are all linear, sub-linear, or log. None are exponential. I don't see strong evidence for doubling in that data set. (Second panel is logistic fit for Jan and Feb p80 numbers.)

6/ Pulling out a few replies with useful followup discussion:
Thanks Basil — that's a great summary of the somewhat confusing state of play. One thing I wonder about are tasks that the agents just can't do well at all. e.g. if 25% of tasks were like that, it would shorten the 80% horizon a lot more than the 50%.

5/ Main conclusions: 1. OpenAI AI R&D time horizon is lower than METR headline... (80% accuracy = ~15min in July) 2. ...but rising v quickly (~2 month doubling time), though this is somewhat sensitive to specification With some possible explanations + additional takeaways above

I'm not sure what to make of that. Perhaps again related to low- vs. high-context human baselines. Or maybe these p50/p80 ratios vary by field substantially...

POINT 4/ the *ratio of p50/p80* time horizons for OpenAI are a lot higher than for METR! With METR, the 50% time horizon was ~4x the 80% time horizon At OpenAI, the 80% time horizon is ~15x the 50% time horizon!

But also! If you measure METR time horizon with "high context" humans instead of the low-context baseliners, then Kwa et al report that time horizon stops being as useful to predict AI capabilities (This is somewhat confusing... 1. The high-to-low-context TH ratio must vary a lot, or systematically, across tasks? While that makes sense, it would be surprising to really destroy any meaningful signal 2. TH measures AI capabilities, but for automating humans, what is *economically* relevant is in fact the high-context capability, not low context... - I should give this more thought and would be especially curious for takes)

Possibility (b): METR measured time horizon using people with "low context" If you measured METR time horizon for people with higher context, time horizons were ~10x lower Perhaps this explains why OpenAI time horizons are so low in levels
In the METR TH paper they report how their TH values are "low context" i.e. they are how long someone would take to do the task if they were not up to speed on the work. In the appendix hey roughly estimates that high context people could be ~10x faster. This doesn't fully explain the gap (since then it would suggest we would be at ~18 minutes 'high context' 80% TH in April, but instead we are only there now) but it makes it less extreme.

Possibility (a): this is just a difficult domain. Like the Kwa cross-domain graph we saw earlier: different intercepts by domain

(GDM's internal AI R&D 80% time horizon is also ~15 minutes, interestingly)
Fortunately, that has arrived. GDM ran a very similar analysis on their internal use of AI for AI R&D and published their graph on page 78 of their Gemini 4 Safety Report:

POINT 3/ Why is the *level* of time horizons for OpenAI internal AI R&D so low? At 80% accuracy, it was only 15-20 minutes as of July. That's not very long!
For me, one of the most interesting parts was pointing out that the current 80% time-horizon for internal OpenAI models when working on real research tasks is only about 15 minutes. That's a *lot* lower than the 80% time horizons one sees from METR's work based on a more generic set of software engineering tasks, or UK AISI's work based on cyber-attack tasks. This performance is weak enough that it means we haven't yet seen evidence of autonomous work on the creative parts of research work (which I think have about a 1 in 3 chance of being a stumbling block for AI being able to match humans across the board).

Explanation (c): time horizon progress has accelerated across domains - this may be consonant with the latest METR report on acceleration at Anthropic... 😐
AECI scores extracted from the graph imply that the time horizon doubling time is now around 2 months, and that the TH of Model 4 is around 240 hours. TBC this is pretty weak evidence. It's unclear if there is a clear 1:1 mapping from AECI to time horizon for numbers this large.