Executive summary: A new report from Ramp, covered by TechCrunch on November 6, 2026, shows AI spending flatlined over the late summer even as actual usage (token volumes) kept rising: token spend grew only 1.5% across August and September, compared with 42% month-over-month in July. Ramp's lead economist, Ara Kharazian, says the slowdown is not a demand problem — it is a price problem. OpenAI halved API prices for its GPT-6 Sol and GPT-6 Luna models in September, and Anthropic followed with significant reductions to Fable 5.1 and Opus 5.5, so the same work now costs less. At the same time, the most experienced buyers are getting thrifty: they route simple tasks to cheap models, set budgets, use defaults that avoid expensive models on long-running jobs, and offload the boring work. The practical upshot for a small business is not "adopt more AI." It is that AI has become a buyer's market, and the opportunity is to buy it the way you'd buy anything else — baseline what you actually spend, route each job to the cheapest model that does it, set budgets and defaults, and offload the cheap-and-boring tasks. The playbook below is four steps you can run this month, with real tools and real pricing, and it ends with a single number — your cost-per-use-case — that decides where to keep spending and where to cut it.
You cannot cut a cost you have not measured. Ramp's data is the reason this step matters now: the market has split into a thin top cohort that hits roughly $10,000 per employee at its summer peak (and has just stuttered a little) and a broad median that spends only $12 to $15 per employee per month, with the top 10% in between at $600 to $1,000. That spread tells you where most small businesses actually are — not at the top, and not at zero, but somewhere in the middle with a few tools you use a lot and a few you use a little. The first move is to make that picture real for your own business. Pull your last 90 days of AI-related spend: the ChatGPT or Claude seats, the API bill if you build anything, the marketing or support tool that quietly includes an AI tier, and any per-seat add-ons. Break it into three buckets — seats (flat monthly), usage (tokens/credits, variable), and tool add-ons (the AI feature inside software you already pay for). The deliverable is not a dashboard; it is one line: "We spend $X/month on AI, of which $Y is flat and $Z is usage." If you cannot answer that question today, that is your problem statement, and it is fixable in an afternoon with your card statements and a spreadsheet. You do not need Ramp or any new tool for this step — you need the last quarter of receipts.
Tools & pricing: Your existing card/spend statements + a spreadsheet (free). If you want automation later, Ramp's spend-management is free for small teams up to a transaction threshold, then priced per transaction; most SMBs can run Step 1 on a spreadsheet at $0.
Here is the "why now," and it is specific and dated. In September 2026, OpenAI cut its API prices in half for GPT-6 Sol and GPT-6 Luna, and Anthropic made significant reductions to Fable 5.1 and Opus 5.5 — and Kharazian's point is that this is now a buyer's market, with Meta and Nvidia also pushing cheap and open models into the mix. The lesson for an SMB is to stop defaulting to the most expensive model for every task and instead build a ladder. At the bottom: a cheap, fast model (think Claude Haiku-class, around $2.50 per million input tokens and $10 per million output tokens) for classification, extraction, routing, and the "is this an emergency?" Yes/No calls that Kharazian specifically calls out. In the middle: a mid model (Sonnet-class, about $10 input / $25 output per million tokens) for drafting, summarizing, and standard multi-step work. At the top: the most capable and expensive model (Opus-class, about $5 input / $25 output per million tokens, with cheaper cached reads) reserved for the genuinely hard, high-stakes jobs where the cost of a wrong answer is real. The rule is: default to the bottom, escalate only when the task genuinely needs it. That single routing habit is most of the cost reduction, because the expensive model stops quietly handling the 80% of tasks that a cheap model was always good enough for.
Tools & pricing (API, per million tokens, input/output): Claude Haiku 5.5 ~$2.50/$10 (cheap tier) · Claude Sonnet 5.5 ~$10/$25 (mid tier) · Claude Opus 5.5 ~$5/$25 (top tier, cached reads cheaper). Consumer plans if you're not using the API: Claude Pro ~$20/mo, Max ~$100/mo, Ultra ~$200/mo; OpenAI's GPT-6 Sol/Luna API prices were halved in September 2026. Verify current rates on the providers' pricing pages before you commit.
The second half of the savings is discipline, and Kharazian's interview is basically a checklist. He says the thrifty, advanced users "use defaults that prevent usage of very expensive models for long-running tasks," "impose some budgets on employees," and "offload some of the spend" to cheaper paths. Translate that into three controls you can set this week. First, a hard per-user or per-project budget with a soft warning at 80% — not a pay-as-you-go shock at month-end. Second, a default model and a default effort: new tasks start on the cheap/mid model and a bounded number of steps, and escalation is a deliberate click, not the default. Third, a length guard: long-running or looping tasks (the ones that burn tokens while you're not watching) get a step cap and a human checkpoint. These three are not about a specific product; they are about making the expensive model opt-in. The realistic effect is that your usage spend ($Z from Step 1) stops compounding on the tasks that do not need it, which is exactly where the gap between "tokens go up" and "spend goes flat" in Ramp's data comes from — the smart buyers stopped paying top dollar for work the cheap model already handled.
Tools & pricing: Budget controls live in your provider console or card-spend tool (Ramp-style per-user caps, or your card's spending rules — often $0 extra). The default-model and step-cap settings are native to Claude/OpenAI developer consoles and to most agentic tools. Cost to add: a couple of hours of configuration, no new subscription.
Once the ladder and the guards are in place, the play flips from cutting to redirecting: point the cheap tier at the high-volume, low-stakes work that used to sit on a person's desk. That is the "offload" Kharazian describes — the email triage, the intake form extraction, the "which vendor is this invoice from?" pass, the first-draft reply, the data entry. Do it on the cheap model with a human skim, and the cost-per-task is pennies. Then close the loop with the one number that makes this a real decision instead a vibe: cost-per-use-case. For each thing you now run on AI, divide that use case's monthly AI cost by the number of times it ran and by the hours it saved. Keep what is cheap-and-high-value, re-tier what is mid-value, and cut what is expensive-and-low-value. Ramp's October data shows a re-acceleration (token spend up 8.9% from September 23 to October 23), which is worth watching but does not change the method: the market is more competitive than it was in July, so the same outcome now costs less, and the businesses that route well, set budgets, and offload the boring work are the ones whose AI line on the P&L stays flat or falls even as they do more with it. That is the buyer's-market outcome, and it is the one you can engineer deliberately.
Tools & pricing: The cheap tier from Step 2 + a two-week measurement (use-case, cost, runs, hours saved). The measurement costs nothing; the spreadsheet is your dashboard.
The shift Ramp's data points to is not "AI got more expensive, so stop." It is the opposite: AI got cheaper, the top cohort is no longer growing the way it was, and the market has tilted into the buyer's favor. For a small business that means the leverage you have right now is cost control — and it is real, not theoretical, because the price cuts (GPT-6 Sol/Luna halved in September, Fable 5.1 and Opus 5.5 reduced by Anthropic) and the cheap/open models from Meta and Nvidia are already live. The gap between a business that burns the expensive model on everything and one that routes each task to the cheapest model that does it, sets a budget, and offloads the boring work is exactly the gap between "tokens go up" and "spend stays flat." The realistic ceiling is not a lower-quality outcome; it is the same outcome at a lower and more predictable cost, so the AI line on your P&L stops being a wild number and becomes something you set. Baseline it this week, ladder the models, make the expensive one opt-in, offload the cheap-and-boring work, and measure cost-per-use-case before you scale. That is the buyer's-market playbook, and the only number that matters is the one you control: cost per unit of useful AI output.
Your next hour should be a baseline, not a strategy debate. Pull your last 90 days of AI spend (seats vs. usage vs. tool add-ons) and write the one line: "We spend $X/month on AI." Then set the model ladder (cheap for triage, mid for drafting, top only for high-stakes work), add a per-user budget and a default-model/step-cap, and offload your top three cheap-and-boring tasks to the cheap tier with a human skim. Measure cost-per-use-case for two weeks, and cut what is expensive-and-low-value. To get the full SMB agentic-AI strategy — including how a model ladder + budgets stacks with the in-suite agents and where the cost control actually pays for itself — start here:
Get the SMB AI adoption strategy