Estimated read time: 7 minutes
Something strange happened to the AI industry this year, and almost none of it reached your bank account. The raw cost of running an AI model has fallen off a cliff. OpenAI’s GPT-5.6 line, generally available since July 9, now bottoms out at roughly a dollar per million input tokens on its economy tier. Anthropic is running introductory pricing on Claude Sonnet 5. Google has been giving away capability that would have been a line item eighteen months ago. Across the board, providers cut 30 to 70 percent in the first half of 2026. Meanwhile the AI features bolted onto your CRM, your scheduling tool, and your email platform cost precisely what they cost in January. Somebody is pocketing the difference, and it isn’t you.
Table of Contents
What Actually Happened to Prices
The mechanism is straightforward competition, arriving all at once. OpenAI broke its single-model pricing convention with GPT-5.6 and shipped three tiers instead: Sol at the top, Terra in the middle, Luna at the floor. The published rates run roughly $5 per million input tokens and $30 per million output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. Terra is positioned as roughly half the cost of the model it replaced with comparable capability, which is the part that should get your attention — not the flashy top tier, but the fact that the middle of the lineup got cheaper without getting worse.
Anthropic, Google, and Meta all moved in the same direction on the same rough timeline. Every major lab now ships an economy model that is genuinely good enough for the overwhelming majority of everyday business tasks: summarizing, drafting, classifying, extracting, answering. The frontier models still matter for hard reasoning and long agentic work. They are no longer required for writing a customer email.
Put a number on it. A small business running fifty million input tokens and ten million output tokens a month — which is a lot of drafting and summarizing, more than most solo operations touch — lands around $110 a month on an economy tier. On a frontier model, the same volume would cost you several times that. Eighteen months ago it would have been a genuine budget conversation. Today it’s a subscription you’d forget you had.
Why Your Bill Stayed Flat
Here’s the uncomfortable structural fact: almost nobody sells you AI. They sell you software that has AI inside it, and the pricing of that software was never indexed to the underlying model cost.
When your project management tool added an AI assistant in 2025, it priced that feature against what customers would pay, not against what it cost to run. That’s normal business. The trouble is that the cost side has since collapsed and the price side has not, because nothing forces it to. Your vendor’s gross margin on the AI feature quietly improved by fifty percent while your invoice stayed identical. There is no line item to point at, no renegotiation trigger, no notification email. It just happened.
This gets more pronounced as embedded agents spread. Gartner projects that 40 percent of enterprise applications will ship with embedded agents by the end of this year, up from under 5 percent in 2025. Every one of those is a vendor buying tokens wholesale and selling outcomes retail. The spread is the business model.
None of this is a scandal. It’s what happens in every industry when input costs fall faster than competitive pressure. It took years for cloud storage price cuts to reach the average SaaS invoice, and in plenty of cases they never did. But it does mean the conventional wisdom — “AI is getting cheaper, so my costs will come down” — is wrong in a specific and expensive way. AI got cheaper for people who buy AI. If you buy software, you’re buying something else.
The Margin Nobody Talks About
Run the arithmetic on a typical AI add-on. A vendor charges $20 per seat per month for an AI writing assistant. A heavy user might generate a few hundred thousand tokens a month through it. At economy-tier rates that’s cents. Even at frontier rates with generous overhead for retrieval, context, and retries, it is very hard to get the true cost above a couple of dollars.
That is not a criticism of the vendor. They built the interface, the integrations, the permissions model, the support desk, and the thing that makes the output actually land in your workflow instead of a chat window. That work has real value and it isn’t free. But you should know the shape of what you’re buying: you are paying for packaging and integration, and the intelligence itself is now close to a rounding error. Which means the right question at renewal is not “is AI worth $20 a seat” but “is this vendor’s packaging worth $20 a seat, given that the ingredient costs almost nothing.”
For some tools the answer is an easy yes. Deep integration into a system where your data already lives is genuinely hard to replicate. For others — the ones where the AI feature is essentially a text box with a prompt behind it — the answer is increasingly no.
Four Things Worth Doing This Month
1. Inventory what you’re paying for AI, separately from software. Most owners can’t answer this. Pull twelve months of statements and tag every line that is an AI add-on, an AI tier upgrade, or a tool you bought specifically for an AI feature. The total is usually larger than the guess, and it’s frequently spread across four or five vendors doing overlapping work. Our SaaS AI risk audit walks through the same exercise from the other direction.
2. Ask for a price adjustment at renewal, explicitly. This works more often than people expect, particularly with mid-sized vendors who know exactly what happened to their input costs and would rather discount than churn you. You don’t need leverage or a threat. “Model costs have dropped substantially this year — is there room on this renewal?” is a complete sentence, and a surprising number of account managers have been authorized to say yes.
3. Consolidate overlapping AI subscriptions. If you’re paying for AI writing in three tools, two of them are redundancy you’re funding out of habit. Pick the one embedded where you actually work and cancel the rest. This is nearly always the single biggest line item recovered.
4. If you use an API directly, move your simple work down a tier. Routing routine requests to an economy model rather than a frontier one commonly cuts a bill 60 to 80 percent with no perceptible quality change. Prompt caching drops repeat-context cost dramatically, and batch processing halves the rate on anything that doesn’t need an answer this second. Most people running API workloads have never touched any of these three levers.
When Building Beats Buying
The falling floor changes the build-versus-buy line, though less than the enthusiasts claim. Two years ago, building an internal AI tool meant a meaningful engineering investment against uncertain model costs. Today the model cost is trivial and the tooling is dramatically better, which moves some jobs across the line.
Where building now wins: narrow, repetitive, high-volume tasks specific to your business. Classifying inbound inquiries by type. Extracting structured fields from the particular invoice format your three biggest suppliers use. Drafting the same category of reply you send forty times a week. These are jobs no vendor will build for you specifically, and they’re now genuinely cheap to run.
Where buying still wins, and will keep winning: anything requiring deep integration with systems you don’t control, anything where you need someone accountable when it breaks, and anything touching customer data where the compliance surface is the actual product. The cost of the model was never the hard part of those.
The honest framing is that cheap tokens made the experiment cheap, not the product. You can now test whether an internal tool would help for a few dollars instead of a few thousand. That’s the real unlock, and most small businesses aren’t using it.
What Happens Next
Two things are likely from here. First, the price floor keeps dropping, because the competitive dynamic that produced this year’s cuts hasn’t resolved — several labs are still buying market share, and economy tiers are the weapon. Second, and more slowly, some of that will finally reach software pricing, but through competition rather than generosity. The vendors who move first will do it as a marketing event, and the ones who don’t will lose customers to a competitor charging less for the same wrapper.
Which means the leverage you have right now is temporary and worth using. Renewals negotiated in the next two quarters are being negotiated against vendors who know their margins improved and don’t yet face pressure to admit it. That asymmetry closes eventually. It’s open now.
FAQ
Are cheaper AI models actually worse?
For most business tasks, no. Economy tiers in 2026 outperform the flagship models of 2024. The gap shows up on genuinely hard reasoning, long multi-step agent runs, and complex code — not on drafting, summarizing, classification, or answering questions from documents you supply. Test your actual workload before assuming you need the expensive tier.
Should I cancel my AI software subscriptions and use the API directly?
Only if you have someone who will maintain it. The API is cheap; the integration, error handling, and ongoing upkeep are not free, they’re just paid in your time instead of dollars. The realistic move for most small businesses is to consolidate subscriptions and negotiate, not to rebuild.
How do I know if my vendor’s AI feature is overpriced?
Ask what it would take to do the same job yourself. If the answer is “paste this into a chat window and read the result,” you’re paying mostly for convenience and should price it accordingly. If the answer involves your customer records, your calendar, and your billing system talking to each other, the integration is the value and the price is probably fair.
Will AI subscription prices come down on their own?
Not without competitive pressure. Vendors reduce prices when a rival forces them to or when churn threatens, not when their input costs improve. Assume you need to ask.
Related Coverage
- SaaS Tools AI Disruption Audit 2026 — which of your subscriptions are most exposed to being replaced outright.
- AI Agent Showdown: What They Actually Do — what the embedded agents showing up in your software are and aren’t capable of.
- Best AI Writing Tools for Small Business — if you’re consolidating, this is the head-to-head on the category people over-subscribe to most.
Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what’s next.