Analysis: AI Cost Ratios Stable Despite Rising Inference Bills
Summary
- • AI tasks cost roughly 3% of equivalent human labor at current frontier models
- • The author argues rising inference costs reflect longer tasks, not higher relative cost
- • Cost ratios have stayed flat across successive frontier models, per this analysis
- • Author contends automation timelines remain on track with METR capability trendlines
Details
Author argues cost ratio has stayed flat, not risen, across frontier models
The author contends that current frontier models complete tasks at their 50% reliability horizon for roughly 3% of the equivalent human labor cost. According to the analysis, this ratio has not increased as model capabilities have improved, meaning affordability has not become a growing constraint alongside capability growth.
Cost ratio defined as AI inference cost divided by human cost for the same task
The author calculates cost ratio by estimating AI inference cost from token counts using OpenRouter pricing, then dividing by estimated human cost per task. All calculations use METR's task weighting methodology, grounding the analysis in publicly available benchmark data rather than proprietary figures.
METR's frontier time horizons are reportedly doubling every few months
METR measures AI capability using 'time horizons' — the length of task an AI can reliably complete. The author references METR's data showing these horizons are growing exponentially. The analysis argues that even when per-task AI spending is capped at 1/32nd of human cost, time horizons still roughly double every three months.
Author contends rising inference bills reflect task length, not worsening cost efficiency
A common interpretation of growing AI compute costs is that automation is becoming economically unviable. The author argues this misreads the data: models are spending more compute because they are completing longer, harder tasks — not because the cost-per-unit-of-work is increasing relative to human alternatives.
Author disputes Toby Ord's conclusion that frontier hourly AI costs are rising exponentially
Researcher Toby Ord previously argued there is moderate evidence that hourly costs at the AI capability frontier are rising exponentially, which would undercut automation economics. The author contends Ord's methodology is flawed and leads to significant overestimates of hourly model cost, dedicating an appendix to rebutting that conclusion.
If the author's thesis holds, automation timelines are not delayed by cost constraints
The author concludes that cost is not an additional bottleneck beyond capability, meaning automation should materialize roughly on the schedule implied by METR's capability trendlines. The analysis also suggests that additional inference-time compute spending would strictly accelerate rather than explain current trends.
Current frontier models complete tasks at ~3% of human labor cost at 50% reliability horizon
This is the central quantitative claim of the analysis, derived from METR task data and OpenRouter pricing. The author argues this figure has remained stable across model generations, which is the empirical basis for rejecting the claim that AI automation is becoming relatively more expensive over time.
Insight = author's analytical argument or interpretation, Research = external benchmark data referenced, Tech Info = methodology or definitional detail, Market Impact = economic or industry implication, Stat = specific quantitative claim
What This Means
This analysis argues that fears about AI automation becoming unaffordable are based on a misreading of inference cost data — and that, in reality, the cost of using AI relative to human labor has stayed roughly constant even as models grow far more capable. If the author's thesis is correct, the main variable governing when widespread AI automation arrives is capability progress, not economics, and METR's trendlines pointing to rapid near-term capability growth should be taken at face value. The piece directly challenges prior work by Toby Ord and reframes rising compute bills — widely interpreted as a warning sign — into evidence that automation is proceeding on schedule. Readers should note this is a single-author analysis piece, not peer-reviewed research, and the conclusions rest on methodological choices that the author acknowledges are contested.