AI Token Costs: Businesses Spend $2,246 Monthly at Median

Actual AI spending patterns revealed

Businesses paid a median of $2,246 per month for AI tokens in April 2026, according to proprietary data from Ramp analyzing thousands of companies' actual vendor payments to providers including Anthropic and OpenAI. However, the average monthly spend reached $140,842—a gap that reflects how a small number of companies with heavy AI adoption or unchecked automated workflows drive costs significantly higher.

The data, first reported by Ramp through Stacker, represents observed effective rates across real usage patterns rather than theoretical pricing from vendor rate cards. Token costs averaged 72 cents per million tokens across all model tiers and usage patterns, but individual model selection creates a 20-fold cost difference: lightweight models like GPT-5-nano run at 7 cents per million tokens, while premium GPT-5.5 costs $1.42 per million tokens.

Why it matters

AI spending has shifted from experimental line items to material budget considerations for many organizations. Understanding actual cost patterns helps finance and technology leaders budget appropriately and identify when AI spend warrants dedicated oversight. The data shows 31% of companies now spend more than $10,000 monthly on AI, while 9% exceed $100,000—thresholds where AI becomes a formal budget line requiring active management.

Per-employee benchmarks show wide variation

Per-employee-per-month (PEPM) spending provides the most useful comparison metric across different company sizes. The overall median stood at $46 PEPM in April 2026, but the middle 50% of companies ranged from $3 to $352 PEPM. Companies using 26 or more different models—a proxy for AI maturity—showed median monthly spending of $26,562.

The median business used nine models during the period, while the average used 16.5. Month-over-month spending swings of 40% or more occurred commonly even with stable headcount, suggesting organizations should build buffer into AI budgets.

What drives costs up and down

Model tier migration represents the single biggest driver of unexpected cost increases. When teams upgrade from lightweight to frontier models for quality improvements, costs can jump 10 to 100 times. Premium models represented 45.8% of tokens consumed but 55.9% of total costs in April 2026, with premium model cost share rising from just 5.7% in June 2025.

Agentic usage—where AI agents autonomously determine how many steps to take to complete tasks—creates unpredictable costs since the agent controls the bill while organizations only approve the initial task. From January 2025 to April 2026, token usage among businesses grew 1,001%, while total spending increased 497%.

Caching provides the most significant cost reduction opportunity. Workflows that reuse the same context achieve 80%-plus cache hit rates, reducing costs by up to five times. Claude Sonnet 4.6 users paid an effective rate of 62 cents per million tokens versus the $3 per million list price, with caching accounting for much of the difference.

Model tier discipline and context management also reduce costs. Nearly one in five Anthropic tokens processed used long-context capabilities, which cost more whether the model needs the full context or not.

The spending data and analysis were produced by Ramp and distributed by Stacker Media.