Every few months a new AI model arrives with a press release about being smarter. The quieter story is that each one also makes last year's smart cheaper. If you budget for AI the way you budget for software licenses, you are probably planning around prices that will not last.
What happened
Two recent launches show the pattern. OpenAI released its GPT-5.6 family in three tiers: Sol, the flagship; Terra, for everyday work; and Luna, the fast, low-cost option. According to MacRumors' report on the June 26 limited preview, OpenAI says Terra performs about like the previous GPT-5.5 at roughly half the price, and Luna sits at its lowest price point. The rollout was also notable because OpenAI agreed to hold back a full public launch while the federal government tested the model, a step the company publicly disagreed with.
Anthropic followed on September 22 with Claude Opus 5.5. The company says it performs at about the level of its more expensive top model on most work, while typical workloads cost about 40% less than the previous Opus 5. List prices dropped from $5 to $4 per million input tokens and from $25 to $20 per million output tokens, and Anthropic says output is more than 30% faster.
The research group Epoch AI measured the long-run trend the same day. In The plunging price of thought, its researchers found that the cost of reaching a given level of AI performance has fallen about 47% per quarter, or roughly 13 times per year, across five benchmarks since 2023. Their example: hitting a set score on a hard science quiz cost about 30 cents per question with an OpenAI model in January 2025. Just under 18 months later, GPT-5.6 Luna matched it for a small fraction of a cent, a 725-fold drop. Epoch calls the pace "faster than for any other transformative technology in history," and it lists caveats: a short three-year dataset, and a real gap between benchmark scores and usefulness.
Why it matters for your business
A token is a chunk of text, usually a word or part of one. You pay per million of them. The numbers above mean three practical things:
- Last year's top model is this year's budget model. Work that once needed the most expensive tier can often run on the middle or cheapest tier now. Many businesses are still paying flagship prices for jobs a smaller model handles fine.
- Tiered families are the new normal. Both OpenAI and Anthropic now sell a ladder of models. The smart move is to put each task on the lowest rung that does it well, not to pick one model for everything.
- Ideas that failed on cost may now pencil out. If you priced an AI project in 2025 and it did not make sense, the math may have changed. Processing every incoming document, call or email used to be expensive. At today's prices, often it is not.
There is a fair counterpoint. Prices fall fastest for capability that already exists. The newest top-end features still cost a premium, and Anthropic's own fast mode for Opus 5.5 costs double the standard rate. Not every bill goes down.
What we'd do
- Review your AI spend every quarter, not once a year. Ask which tasks could move to a cheaper tier.
- Keep a small test set of 20 to 50 real examples from your business. When a new model or tier comes out, run them through it and compare quality and cost in an afternoon.
- Avoid long contracts priced on today's rates unless they include a way to switch models or reprice.
- Revisit shelved projects. Pull out anything you rejected on cost in the last 18 months and rerun the numbers.
- Do not chase every launch. A new model is only worth switching to if it is better or cheaper on your own test set.
Our take
We think the most important AI number for a business owner is not a benchmark score. It is the price of the capability you already need, and that number keeps dropping. Businesses that build a simple habit of testing and switching will keep getting more for less. Businesses that lock in one model at one price will quietly overpay. The launches get the headlines. The price curve is what changes your margins.
Want a system like this in your business?
A discovery conversation is where we map it: what to automate first, where the human checkpoints go, and what it would take.
Book a discovery conversation →