Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents (Nov 23, 18:30 MSK): 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →

US AI Got 1.91x Better, but Costs Rose 6.19x: Return per Dollar Fell 69%

A METR report compared what American AI achieved after October 2025 and what it cost. Model capabilities rose 1.91x, while R&D spending rose 6.19x, like a factory that nearly doubled output on a sixfold budget. The gain per dollar fell about 69%, and the gain per technical employee per year fell 38%. Planning a product around cheap jumps in model quality is still risky.

AI-processed from TG: @bezsmuzi (Русский ИТ бизнес); edited by Hamidun News
US AI Got 1.91x Better, but Costs Rose 6.19x: Return per Dollar Fell 69%
Source: TG: @bezsmuzi (Русский ИТ бизнес). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @bezsmuzi (Русский ИТ бизнес) ↗
Editor-in-Chief
AI capabilities×1.91, R&D spending: ×6.19, headcount: ×3.10
Key metric #2Return per dollar fell 69%, per employee per year 38%
Key metric #3Economists Garicano and Imas see no evidence of AI self-acceleration

The common assumption is that AI is about to start improving itself: writing code for the next model, running experiments, and progress snowballing. A new report from the research organization METR paints a more down-to-earth picture. Since October 2025, American AI has indeed gotten better, but each step forward has become noticeably more expensive.

What METR Calculated and How to Read These Numbers

The idea is simple: compare what the labs got with what they spent to get it. R&D, or research and development, covers everything companies invest in building new models: compute, data, salaries.

After October 2025:

  • measured AI capabilities rose 1.91x;
  • R&D spending rose 6.19x;
  • technical headcount rose 3.10x.

Dividing the result by the inputs gives: the capability gain per dollar invested fell about 69%, and the gain per technical employee per year fell 38%. This is easy to check: 1.91 / 6.19 ≈ 0.31, meaning the return is about a third of what it was. And 1.91 / 3.10 ≈ 0.62, meaning a decline of about 38%.

An example in plain terms. Imagine a factory that nearly doubled its output, but to do it, it had to increase its budget sixfold and hire three times as many people. There is more product, but every ruble invested works worse than before. That is roughly the picture the report paints for American labs.

Why This Matters for Talk About "Self-Accelerating"

AI

In theory, the opposite should happen. If models help researchers, progress per unit of cost should rise: AI speeds up research, research makes AI better, and better AI speeds research up even more. Forecasts of an "explosive" takeoff are built on this logic.

Economists Luis Garicano and Alex Imas do not yet see convincing evidence in the available data that AI itself generates the resources for research. And that is precisely the condition explosive growth would require. For now, progress has to be paid for with money, compute, and people, and those costs keep climbing.

An important caveat: METR's numbers do not say that progress has stopped. Capabilities nearly doubled. The point is that each improvement now takes disproportionately more resources.

What Developers, Product Managers, and Founders Should Do

First, do not plan on models getting endlessly cheaper and improving in leaps. If the cost of an improvement grows faster than the improvement itself, labs will recover their costs through prices, rate limits, and a focus on profitable segments. Build a possible rise in API costs into your budget.

Second, build an architecture that is not tied to a single model. If the next version turns out to be less of an improvement than promised, the winners will be those who know how to squeeze the most out of current models: good prompts, quality evaluation on their own data, careful context management, and the right tools.

Third, read loud forecasts with caution. The phrase "AI will accelerate itself" remains a hypothesis, not a measured fact.

Limitations

The original post is a brief retelling of the report. It contains no details on how METR measures "capabilities," which models and labs were included in the calculation, or how R&D spending was counted. Those choices affect the final percentages, and a single report covering one period does not prove a long-term trend. The conclusion concerns the ratio of result to investment, not the absolute speed at which new models appear.

It therefore makes sense to treat the report as a warning signal and wait for confirmation in subsequent periods. If that confirmation comes, the conversation about AI will need to shift from fantasies of avalanche-like growth to economics: who is willing to pay, and how much, for each next step.

Story

Key Questions Answered

What METR Calculated and How to Read These Numbers?

The idea is simple: compare what the labs got with what they spent to get it. R&D, or research and development, covers everything companies invest in building new models: compute, data, salaries. After October 2025: measured AI capabilities rose 1.91x; R&D spending rose 6.19x; technical headcount rose 3.10x. Dividing the result by the inputs gives: the capability gain per dollar invested fell about 69%, and the gain per technical employee per year fell 38%.

Why This Matters for Talk About "Self-Accelerating" AI?

In theory, the opposite should happen. If models help researchers, progress per unit of cost should rise: AI speeds up research, research makes AI better, and better AI speeds research up even more. Forecasts of an "explosive" takeoff are built on this logic. Economists Luis Garicano and Alex Imas do not yet see convincing evidence in the available data that AI itself generates the resources for research. And that is precisely the condition explosive growth would require.

What Developers, Product Managers, and Founders Should Do?

First, do not plan on models getting endlessly cheaper and improving in leaps. If the cost of an improvement grows faster than the improvement itself, labs will recover their costs through prices, rate limits, and a focus on profitable segments. Build a possible rise in API costs into your budget. Second, build an architecture that is not tied to a single model.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…