American enterprises are integrating artificial intelligence into their day-to-day operations at record volumes, yet the financial cost of running those workloads is steadily contracting. According to the latest Ramp AI Index—which analyzes aggregated transaction data and telemetry across more than 70,000 corporate clients using the Ramp finance automation and card platform—median per-employee AI expenditure among the top 1% of enterprise spenders declined by 9.7% in August 2026, falling to $7,205 per employee per month from an all-time peak of approximately $8,000 in July.
The spending contraction is not an indicator of waning enterprise interest; rather, it demonstrates that corporate AI adoption is undergoing a maturation cycle. As foundation model developers engage in intense price competition, the blended effective price per 1 million tokens has collapsed by 41% from its March 2026 highs down to $0.68. Confronted with cheaper compute and sophisticated internal spend routers, enterprise engineering and finance teams are actively “trading down”—shifting routine production workloads away from expensive, frontier flagship architectures toward cost-effective standard and lightweight models.
Key Takeaways
- Top 1% Cut Spend by 9.7%: Median per-employee monthly AI spend among the highest-spending 1% of US firms dropped from an ~$8,000 peak in July to $7,205 in August, despite broad-based adoption continuing to expand.
- 41% Token Price Deflation: The volume-weighted effective cost per million tokens across input, output, and cached prompts plunged 41% between March and September 2026, reaching a new low of $0.68.
- The “Trading Down” Migration: Frontier models (such as Claude Opus and OpenAI’s GPT-5.6 Sol) saw their collective token share contract from 53% in early August to 45% by early September, as enterprise workloads migrated toward standard tiers like Claude Sonnet and GPT-5.6 Terra.
- Anthropic Retains Enterprise Lead: In August, 43.8% of US businesses on Ramp transacted with Anthropic (up 0.34 percentage points), maintaining an edge over OpenAI at 39.8% (up 0.09 percentage points).
- Open-Source Stalls in Enterprise: Despite developer enthusiasm, open-weight and self-hosted models represent only 6.4% of AI-spending companies and just 3.6% of all tracked enterprises, highlighting the enduring convenience of managed commercial APIs.
- The Revenue Question: Ramp Chief Economist Ara Kharazian cautioned that while query volume is expanding exponentially, it remains uncertain whether token growth will expand quickly enough to offset continuous unit-price deflation for frontier AI labs.
The Strategic Turning Point: From Frontier Exuberance to Portfolio Routing
For the past two years, enterprise AI adoption was characterized by indiscriminate procurement: companies bought access to the most powerful, cutting-edge foundation models available, often applying expensive frontier models to simple tasks like customer email drafting, data extraction, and routine report summaries.
The August data from Ramp confirms that enterprise procurement has shifted toward cost-calibrated architectural routing:
THE ENTERPRISE COST OPTIMIZATION SHIFT
│
┌────────────────────────────────┴────────────────────────────────┐
▼ ▼
THE "FRONTIER ONLY" ERA (2024–Early 2026) THE "MODEL PORTFOLIO" ERA (Late 2026)
• Single flagship model for every task • Multi-tiered model routing architectures
• Unmonitored token usage and high API bills • Algorithmic task-to-tier matching
• Frontier model token share: >53% • Frontier token share drops to 45%
• Average blended token price: $1.15–$1.45/M • Blended token price falls 41% to $0.68/M
│ │
└────────────────────────────────┬────────────────────────────────┘
▼
THE ENTERPRISE REALITY
Workload volume is rising double-digits,
while per-employee financial costs compress.
- Enterprise FinOps Discipline: Corporate finance and IT departments have deployed Token Spend Management (TSM) systems to set department-level token budgets, monitor prompt caching efficiency, and enforce rate limits. Major corporations including Uber, Amazon, and Walmart have placed strict token limits and model guidelines on internal employee seats.
- Algorithmic Model Cascading: Rather than sending every prompt to top-tier reasoning engines, developers are implementing gateway routers that direct 80% of routine queries to high-throughput, low-cost models, reserving flagship models solely for multi-step reasoning, mathematical theorem proving, and complex code refactoring.
- Prompt Caching and Architectural Efficiency: Widespread adoption of prompt caching across both Anthropic and OpenAI APIs has reduced the cost of passing large system prompts and background documentation, drastically cutting repetitive input token charges.
Dissecting the Metrics: What the Ramp AI Index Shows
Ramp’s dataset—drawn from corporate card transactions, ACH disbursements, invoice payments, and direct API spend telemetry across 70,000+ businesses—provides a granular look at the shifting economics of business AI:
+-----------------------------------------------------------------------------------+
| ENTERPRISE AI ADOPTION & UNIT ECONOMICS (RAMP AI INDEX) |
+-----------------------------------------------------------------------------------+
| Operational / Spending Metric | Reported Value / Trajectory (August–Sept 2026) |
+--------------------------------+---------------------------------------------------+
| **Overall Business AI Adoption**| **56.1% of all US businesses on Ramp** |
| **Census Bureau Comparison** | ~22.1% (Undercounts unmonitored employee tools) |
| **Top 1% Spend Per Employee** | **$7,205 / month** (Down 9.7% from July peak) |
| **Top 10% Spend Per Employee** | **~$630 / month** (Continuing modest climb) |
| **Median Spend Per Employee** | **~$12 / month** (Stable, gradual baseline rise) |
| **Blended Price Per 1M Tokens**| **$0.68** (41% drop from March 2026 peak) |
| **Frontier Token Volume Share**| **Fell from 53% to 45%** (Opus, Sol, Mythos tier) |
| **Anthropic Corporate Adoption**| **43.8% of companies** (Market share leader) |
| **OpenAI Corporate Adoption** | **39.8% of companies** (Runner-up) |
| **Open-Weight Model Adoption** | **6.4% of AI-paying firms** (3.6% of all firms) |
+--------------------------------+---------------------------------------------------+
1. The Spend Divergence Across Percentiles
The 9.7% spending decline was concentrated in the top 1% cohort—consisting of high-growth technology startups, quantitative hedge funds, and engineering-heavy enterprises that spend hundreds of thousands of dollars monthly on AI APIs.
In contrast:
- The top 10% cohort (spending ~$630/employee/month) and the median cohort (~$12/employee/month) continued to see slight sequential spending increases.
- This divergence indicates that while mainstream enterprises are gradually rolling out basic software seats, the most sophisticated power users have completed their initial exploratory phase and are actively squeezing waste out of their API integrations.
2. The 41% Collapse in Token Pricing
The volume-weighted price per million tokens dropped from peaks above $1.15–$1.40 in early spring 2026 to $0.68 by late summer.
- Model providers have slashed prices on legacy and standard tiers to maintain developer retention.
- OpenAI has maintained an aggressive pricing posture, pricing standard tokens lower than Anthropic across comparable context windows.
- Aggressive prompt caching discounts—frequently cutting cached read costs by 75% to 90%—have lowered the effective per-token rate paid by production workloads.
Market Share: Anthropic Maintains Lead Over OpenAI as Open Weights Lag
The Ramp AI Index details the ongoing commercial duel between San Francisco’s leading foundation model providers, as well as the commercial standing of open-source software:
ENTERPRISE ADOPTION FOOTPRINT
│
┌──────────────────────────────────┼──────────────────────────────────┐
▼ ▼ ▼
ANTHROPIC (CLAUDE ECOSYSTEM) OPENAI (GPT ECOSYSTEM) OPEN-WEIGHT ALTERNATIVES
• 43.8% enterprise penetration • 39.8% enterprise penetration • 6.4% of AI spenders (3.6% total)
• Anchored by Claude Code & • Gained momentum with GPT-5.6 Sol • Self-hosting complexity & GPU
engineering agent workflows and low-cost standard models overhead limit enterprise uptake
Anthropic’s Software Development Moat
Anthropic maintained its position as the most widely adopted enterprise provider on Ramp, with 43.8% of companies paying for its services compared to OpenAI’s 39.8%.
Anthropic’s edge is sustained by its developer tooling:
- Products like Claude Code and the Sonnet tier have established a dominant footprint among engineering teams building autonomous coding agents, refactoring legacy repositories, and managing automated CI/CD pipelines.
- However, month-over-month growth slowed to 0.34 percentage points for Anthropic and 0.09 percentage points for OpenAI, suggesting enterprise penetration in tech-adjacent sectors is approaching initial saturation.
The Open-Source Paradox
Despite the widespread public visibility of open foundation models (such as Meta’s Llama family, Mistral, and DeepSeek), open-weight architectures have made minimal headway in enterprise billing:
- Only 6.4% of AI-using companies on Ramp pay for infrastructure specifically running open-source models.
- Across all 70,000+ businesses, open-source adoption stands at just 3.6%.
- For most enterprise teams, the overhead of provisioning dedicated cloud GPU instances, managing low-level inference runtimes (like vLLM or TensorRT-LLM), and ensuring high availability outweighs the unit-cost savings of open weights, keeping enterprises reliant on managed commercial APIs.
The Jevons Paradox Paradox: Can Volume Outpace Price Compression?
The central economic question raised by Ramp Chief Economist Ara Kharazian is whether the AI sector is experiencing a true Jevons Paradox—an economic phenomenon where increased efficiency leads to lower unit costs, which in turn sparks an exponential explosion in demand that increases total consumption expenditure.
THE AI ECONOMICS SCENARIO
│
┌──────────────────────────────┴──────────────────────────────┐
▼ ▼
SCENARIO A: CLASSIC JEVONS PARADOX SCENARIO B: DEFLATIONARY COMPRESSION
• Token cost drops 41% • Token cost drops 41%
• Businesses build 10x more agents • Workloads stabilize on cheaper models
• Net enterprise AI spend expands • Token volume grows, but cannot offset cuts
• Foundation lab revenues surge • Lab margins face structural squeeze
Historically, software compute followed Jevons Paradox: as cloud server storage and compute costs fell, companies built vastly larger applications, swelling AWS and Microsoft Azure revenues.
In the enterprise AI market of late 2026, the equation is less certain:
- While companies are deploying autonomous agents and processing larger context windows, engineering teams are learning to make each workflow leaner.
- Techniques like speculative decoding, smaller embedding models, and synthetic dataset distillation allow companies to accomplish complex tasks with fewer total prompt calls.
- If token demand growth fails to accelerate past the 40%+ annual decline in token prices, foundation model providers could face margin compression, even as the utility delivered to the broader economy reaches all-time highs.
What Could Happen Next?
- Anthropic’s Public Filing: Market attention will turn toward Anthropic’s anticipated initial public offering (IPO) filing, where audited financial disclosures will reveal whether high enterprise adoption translated into sustainable gross margins amid token deflation.
- Aggressive Tier Bundling: To protect average revenue per user (ARPU), providers like OpenAI and Google will accelerate enterprise subscription bundling, pairing model tokens with enterprise collaboration software, automated security screening, and proprietary database connectors.
- The Rise of Edge and On-Device Offloading: As smaller 2B–8B parameter models gain competence, enterprises will begin shifting initial classification and data-masking steps entirely on-device, applying further downward pressure on centralized cloud API expenditures.
Frequently Asked Questions (FAQs)
What does the Ramp AI Index show regarding corporate AI spending?
The Ramp AI Index shows that while enterprise AI adoption continues to rise, businesses are spending less on a per-employee basis. In August 2026, the top 1% of enterprise AI spenders cut their monthly per-employee AI costs by 9.7% to $7,205, driven by falling token prices and a shift toward cheaper model tiers.
Why have AI token prices fallen so dramatically?
The blended effective price per million tokens fell 41% from its March 2026 peak to $0.68. This deflation is driven by aggressive price competition among foundation model providers (such as OpenAI, Anthropic, and Google), hardware efficiency gains, architectural improvements like prompt caching, and developer routing to smaller, specialized models.
What is meant by enterprises “trading down” their AI models?
“Trading down” refers to companies shifting routine workloads away from expensive, top-tier “frontier” reasoning models (like Claude Opus or GPT-5.6 Sol) toward “standard” or “lite” models (like Claude Sonnet or GPT-5.6 Terra). The token share of frontier models dropped from 53% to 45% as businesses realized cheaper models could handle the majority of tasks effectively.
Who leads the enterprise market between OpenAI and Anthropic?
According to Ramp’s August transaction data, Anthropic leads in corporate penetration with 43.8% of tracked US businesses paying for its services, while OpenAI accounts for 39.8%. Anthropic’s lead is largely driven by strong adoption among software engineering teams using tools like Claude Code.
Are businesses switching to open-source AI models?
Not significantly. Despite high interest among independent developers, only 6.4% of AI-paying enterprises on Ramp and 3.6% of all companies run open-weight models. Most businesses still favor the reliability, compliance guarantees, and zero-maintenance infrastructure of managed commercial APIs.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
