AI prices are falling. So why are AI bills going up?

If you pay for AI by usage, the past few weeks brought good news. OpenAI and Anthropic both cut prices on their newest models, some by half or more. Yet across thousands of businesses, total spending on AI keeps climbing.

Both things are true at once, and understanding why is the key to budgeting for AI without surprises.

The price cuts

• OpenAI, September 22. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, half the $4 and $20 of GPT-5.6 Sol. The smaller GPT-6 Luna costs $0.10 and $0.50. OpenAI told VentureBeat these are “permanent prices, not promotional or introductory pricing.”
• Anthropic, October 7. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts of up to 100,000 tokens, down from $1 and $5 for Haiku 4.5. Anthropic says it costs “around 75% less to run” on average.
• What businesses actually pay. Ramp, which processes AI payments for thousands of companies, reported on September 9 that the effective price had fallen to $0.68 per million tokens, 41% below March’s peak of $1.15.

A token is a piece of a word. Models charge for the tokens they read and the tokens they write.

And yet, the bills are climbing

Ramp’s data shows that token usage among businesses grew 1,001% from January 2025 to April 2026, while spending grew 497%. Prices fell, but usage grew far faster. The median business on Ramp spent $2,246 a month on AI in April. Ramp’s summary: “AI agents run like a meter.”

Two things drive that.

Agents use far more tokens than chatbots. A chatbot reads your question and writes an answer. An agent plans, calls tools, reads the results, rereads documents, checks its work, and tries again. Every step sends text back and forth, and you pay for all of it. Futurum Research estimates that agentic AI “can increase token consumption per task by 10 to 100 times” compared with a simple request. That research was sponsored by QumulusAI, which sells computing capacity, so read the upper end with that in mind.

Price per token isn’t cost per task. Newer models can count the same text differently. Anthropic’s documentation notes that Haiku 5.5 counts the same input text as approximately 30% more tokens than Haiku 4.5 did. The price cut is still large, but the only number that matters is what a finished task costs.

Five ways to keep AI costs predictable

1. Budget per task, not per token. Measure what one finished job costs: one invoice processed, one call answered, one report written. That’s the number to compare with the time it saves.
2. Use the smallest model that does the job. Send routine work to fast, inexpensive models, and save the expensive ones for the hard cases.
3. Set limits and alerts. Most providers let you cap spending or warn you as it climbs. Turn that on before an agent goes live, not after the first surprise.
4. Don’t pay to reread. Agents that resend the same long documents at every step waste tokens. Good design summarizes, caches, and fetches only what’s needed.
5. Price out the steady work. If AI runs all day on predictable, high-volume tasks, compare the meter with owning the hardware. Private AI on your own machine is mostly an upfront cost, and heavy use doesn’t run up a bill.

When on-prem wins, and when it doesn’t

Owning the hardware isn’t automatically cheaper. If your AI use is light or comes in bursts, cloud AI is hard to beat, especially at today’s prices. On-prem makes sense when use is heavy and steady, when data needs to stay in the building, or when you want a fixed cost you can plan around. Hardware has its own cost pressures right now, which we covered in Private AI now fits on one desktop.

The right answer is often both: private AI for the steady, sensitive work, and a cloud model for the occasional heavy lift. We help businesses find that split based on how they actually use AI.

Sources

• VentureBeat: OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more, September 22, 2026
• Anthropic: Claude Haiku 5.5, October 7, 2026
• Claude documentation: What’s new in Claude Haiku 5.5
• Ramp: AI token costs for businesses, June 8, 2026
• Ramp: AI Index, September 2026, September 9, 2026
• Futurum Research, via Business Wire: Agentic AI can raise token use per task up to 100 times, September 25, 2026