AI pricing per token is falling, and Ramp’s spend data shows US businesses paid 41% less per million tokens in September 2026 than in March. Bills still climb because cheaper AI gets used more, and vendors keep changing the unit they bill by, so budget each AI renewal on the billing unit and expected volume.
Key Takeaways
- AI pricing per token fell 41% from March to September 2026, but cheaper AI gets used more, so total spend can still climb.
- Your AI bill depends on 2 numbers you can forecast: the vendor's billing unit and how much of it your teams use.
- Vendors change billing units often, so ask whether your contract lets them change what a credit buys during your term.
- Year-one usage is hard to predict, so budget it as a ramp and ask each vendor what counts against your allowance.
Why is your AI bill going up when AI is getting cheaper?
In May I wrote here that AI prices were going up and would keep going up. On the price per unit of AI, the 2026 numbers say otherwise. AI models charge by the token, a chunk of text about the size of a short word. Ramp, a corporate card and spend management platform, tracks what US businesses pay per million tokens from its customers’ transactions. That price fell 41% between March and September 2026, from $1.15 to $0.68, as OpenAI and Anthropic cut prices and companies moved routine work to cheaper models (1. Kharazian, 2026).
Cheaper AI gets used more, and that has outrun price cuts before. Across the whole market, Menlo Ventures, the venture capital firm, estimated that business spending on AI models more than doubled, from $3.5 billion to $8.4 billion, between late 2024 and mid-2025, while the price per token kept dropping (2. Menlo Ventures, 2025). Inside one company, cheaper AI means teams push more tasks through it, wire it into more workflows, and let agents run longer chains of steps.
And the best models stay expensive. Gorkem Yurtseven, co-founder of fal.ai, a company that sells access to AI models, told Bessemer Venture Partners that “everyone wants the best model, and those are more expensive to run” (3. Bessemer Venture Partners, 2026). Your team will probably reach for the best model whenever the work matters.
The case for rising prices hasn’t gone away. Bessemer, a venture firm that invests in AI companies, puts their gross margins well below traditional software’s, because every AI request costs the vendor money to run (3. Bessemer Venture Partners, 2026). So far, the vendors in this piece have answered that pressure by changing how they bill. Your renewal quote lists the vendor’s own units, like credits, seats, or conversations. Ramp’s 41% drop covers what companies pay the AI model makers directly, so it can’t tell you whether your vendor’s price per credit or seat is going down. So a 2027 AI budget built on a price-change percentage, up or down, forecasts the wrong number. Forecast these 3 things, then fix the contract.
What unit are you paying for?
Vendors bill you by whatever unit they choose, and they keep changing it.
Salesforce launched its Agentforce agents at $2 per conversation, whatever the conversation involved. In May 2025 it added Flex Credits, which charge $0.10 for each action the agent takes, like looking up a record or booking a meeting. Its own example prices the same sales exchange at $0.30 to $0.60 under the new model. It also added a flat per-user license with unlimited use (4. Salesforce, 2025). That’s 3 ways to pay for the same agent inside a year.
Salesforce’s example made that exchange cheaper. But per-action billing also means cost now grows with the number of steps, and Salesforce says so directly: “simple tasks cost less and complex tasks cost more” (4. Salesforce, 2025). Each change rewrote what an invoice counts and how fast the count grows, which is why last year’s bill is a poor guide to next year’s. Before you budget, write down the unit for every AI tool you pay for: seats, credits, actions, conversations, resolutions, or tokens. If your team can’t name the unit, finance can’t forecast it.
How much will your teams use?
The unit tells you what gets counted. Volume tells you how many. Pull the last 3 to 6 months of usage from each vendor’s admin console or invoices, and look at the trend line. An average hides the climb.
Plan for usage to keep moving after rollout. More people adopt the tool, teams connect it to more workflows, and an agent that started on one support channel ends up on all of them. Budget year one as a ramp, and ask each vendor which activity counts against your allowance, including the test runs your team does while tuning an agent.
Can the vendor change the unit during your term?
AI vendors are still working out how to charge. HubSpot moved 2 of its AI agents from per-conversation and per-contact charges to per-result pricing in April 2026 (6. HubSpot, 2026). ICONIQ, a growth investment firm, found that a sizable minority of companies building AI products plan to change their pricing model again within a year (7. ICONIQ, 2026). If your contract lets the vendor change what a credit buys, your budget moves whenever their pricing does.
Read the renewal for 3 things: whether the vendor can change how many credits a task costs, whether it can change the price per unit mid-term (the rate card), and how much notice you get. Then ask for what you need. A locked rate card for the term is the first ask. Notice of at least one full budget cycle before any change to the unit is the second.
What should go in the contract?
3 contract terms limit what a unit change can do to your bill. Stacia Garr, co-founder of RedThread Research, a firm that studies HR and people analytics technology, worked through per-unit AI pricing in Harvard Business Review. She shows what a small change in the unit rate does to the bill: “What if a vendor increases its unit price from one cent to two? The identical usage could double in cost” (5. Garr, 2026). She recommends caps that stop runaway billing, a grace period before any price change, and rights to roll unused credits into the next year so they don’t expire on January 1.
Add one more: a usage report you can read. If the vendor’s console can’t show what each team and each use case consumed, you won’t know which workflow is driving the bill until the invoice arrives.
An unlimited per-user license, like the one Salesforce now offers, removes the forecasting work, and you pay for it with a commitment to every seat, used or not. For teams with steady, heavy use, it’s worth it. For everyone else, a capped usage plan is easier to keep in check.
Where to start before the budget locks
Pick your 3 largest AI line items. For each one, write down the billing unit, last quarter’s volume, and whether the contract lets the vendor change either. That fits on a single page, and it’s the page you bring to the renewal.
Before you approve next year’s AI line, ask each vendor one question: what unit am I paying for, and can you change it during my term?
Frequently Asked Questions
Is AI pricing going up or down in 2026?
What is credit-based pricing for AI tools?
Why is my AI bill going up if AI is getting cheaper?
How should I budget for AI software in 2027?
Will AI vendors change their pricing models again?
Is an unlimited AI license safer than usage-based pricing?
References
- Kharazian, A. (2026). Cracks in the AI thesis part 2. Ramp Economics Lab. https://econlab.substack.com/p/ai-index-sept-2026
- Menlo Ventures. (2025). 2025 mid-year LLM market update: Foundation model landscape and economics. Menlo Ventures. https://menlovc.com/perspective/2025-mid-year-llm-market-update
- Bessemer Venture Partners. (2026). The AI pricing and monetization playbook. Atlas. https://www.bvp.com/atlas/the-ai-pricing-and-monetization-playbook
- Salesforce. (2025). Flexible pricing is a full circle moment for Salesforce. Salesforce. https://www.salesforce.com/blog/flex-credits
- Garr, S. (2026). How to respond to the coming AI cost shock. Harvard Business Review. https://hbr.org/2026/08/how-to-respond-to-the-coming-ai-cost-shock
- HubSpot. (2026). HubSpot’s Customer Agent and Prospecting Agent: Now you pay when the task is complete. HubSpot. https://www.hubspot.com/company-news/hubspots-customer-agent-and-prospecting-agent-now-you-pay-when-the-task-is-complete
- ICONIQ. (2026). State of AI: Bi-annual snapshot. ICONIQ. https://www.iconiq.com/growth/reports/2026-state-of-ai-bi-annual-snapshot
