Top AI spenders cut per-employee costs by nearly 10 percent in August
Median per-employee AI spending among the top 1% of US companies fell 9.7% in August to $7,205, marking the first notable decline among the highest-spending firms The effective price per million tokens dropped 41% from its March 2026 peak to $0.68, driven by aggressive price cuts from OpenAI and Anthropic Frontier model token share declined from 53% in early August to 45% by early September as companies migrated toward cheaper standard models like GPT-5.6 Terra and Claude Sonnet Open-weight mode
Analysis
TL;DR
- Median per-employee AI spending among the top 1% of US companies fell 9.7% in August to $7,205, marking the first notable decline among the highest-spending firms
- The effective price per million tokens dropped 41% from its March 2026 peak to $0.68, driven by aggressive price cuts from OpenAI and Anthropic
- Frontier model token share declined from 53% in early August to 45% by early September as companies migrated toward cheaper standard models like GPT-5.6 Terra and Claude Sonnet
- Open-weight model adoption remains minimal, with only 6.4% of AI-using companies on Ramp's platform running open-weight models
- Ramp's chief economist flagged "cracks in the AI thesis," citing weak frontier model adoption and the cost-driven migration as warning signs for model providers
Why It Matters
This data reveals a critical inflection point in enterprise AI economics: the era of unlimited spending on frontier models is showing signs of fatigue as companies prioritize cost efficiency without sacrificing adequate performance. For AI practitioners and providers, the shift signals that the market is maturing beyond the initial hype cycle, with procurement policies increasingly favoring "good enough" models over the most expensive options.
Technical Details
- Ramp's AI Index tracks monthly US business spending on AI services; the September 2026 edition shows Anthropic adoption at 43.8% of US companies (up 0.34 pp) and OpenAI at 39.8% (up 0.09 pp), with IT and finance sectors leading adoption
- Token pricing collapsed 41% from March 2026 peak to $0.68 per million tokens, with OpenAI consistently priced lower than Anthropic across comparable offerings
- Frontier models (Opus, Fable, Sol) lost 8 percentage points of token share in a single month, while standard models (GPT-5.6 Terra, Claude Sonnet series) absorbed the volume growth
- Open-weight model adoption is negligible: 6.4% among AI-using companies on Ramp and 3.6% across all companies, with actual adoption likely even lower given Ramp's routing platform structure
- The top 1% spending cohort is small and volatile, with the report noting estimates may be revised; seasonal factors (August vacations) account for part of the decline
Industry Insight
- Model providers should expect continued margin pressure as enterprise buyers institutionalize cost-aware AI procurement policies; the "trade-down" trend from frontier to standard models is likely structural, not temporary
- The weak adoption of Fable 5 and frontier models suggests that raw capability gains face diminishing returns in enterprise contexts, where reliability, cost, and integration matter more than peak performance
- Open-weight models remain a non-factor in enterprise adoption despite their cost advantages; providers should not count on open-source disruption in the near term, but the gap may narrow as open models improve in quality
Disclaimer: The above content is generated by AI and is for reference only.