A New Model Showed Up, and Nobody Read the Fine Print
Anthropic released a new top-tier model this week. Somewhere on a retail team, someone saw it sitting at the top of the model picker, switched over to try it, and just left it there.
Nobody told them that the moment a top-tier model isn’t included in your plan, it starts billing on pay-as-you-go credits, often at several times the rate of whatever model they were using the day before. Why would they know that? It’s just a dropdown.
By day two of the billing cycle, that retailer had spent their entire month’s AI budget.
So here’s what’s actually going on underneath that: what you’re paying for, what the price gap between models really looks like, and the handful of habits that stop this happening to you.
What You’re Actually Paying For
Everything an AI tool does gets measured in tokens, small chunks of text. You’re charged for what goes in (your prompt, any documents, the whole conversation so far) and for what comes out. Output almost always costs more than input.
If your team is on a standard monthly subscription, you never actually see a token bill. Token usage just eats into a rolling time-based allowance in the background. But the moment you move to pay-as-you-go credits, or you’re calling a model directly through a developer platform, every single token shows up on the invoice.
That distinction is exactly where most budget surprises come from.
The Real Cost Gap Between Models
Anthropic’s Claude models sit at very different price points. The exact dollar figures move around, so here’s the gap that actually matters for your budget: how much more expensive each tier is relative to the cheapest one.
| Model | Relative Output Cost | Best For |
|---|---|---|
| Haiku 4.5 | Baseline | Quick, high-volume jobs |
| Sonnet 5 | ~2x baseline | Most day-to-day retail work |
| Opus 5 | ~5x baseline | Complex analysis, harder problems |
| Fable 5.1 | ~10x baseline | Long, multi-step agent workflows |
Same product description, same email, same spreadsheet. Run it through the wrong model and you can pay ten times more for the exact same output.
Your Subscription Doesn’t Automatically Include Everything
A standard paid seat gives you a proper time-based allowance, and the mid-tier and entry-level models sit inside it. The top-tier model usually doesn’t. It sits right there in the model picker looking like just another option, but every message sent to it runs on separate pay-as-you-go credits. No credits enabled, and it simply won’t run. Credits on, and it runs and bills.
A premium seat typically does include the top-tier model, but only up to a portion of your weekly allowance. Once that portion’s used up, you’re back to switching models or paying on top.
The real question isn’t whether the newest model is better. For a handful of hard, long-running tasks, it genuinely is. The question is which task in your business is actually worth paying several times more to solve. For most retail teams, the honest answer is one or two tasks, and they’re rarely the ones running before nine on a Tuesday.
The Same Trap Shows Up on ChatGPT Too
If your team runs on ChatGPT instead of Claude, the mechanics look different but the shape is identical. The base plan doesn’t include the top reasoning models at all. The premium tiers get access up to a weekly message allowance, and once that’s used up, the assistant quietly drops back to a smaller model, or locks the bigger one out until the counter resets.
Nobody tells the person at the desk when that happens. They just notice the output getting worse that afternoon and assume something broke.
Different vendor, same fix. Don’t build your habits around a model name, they get renamed and replaced every few months. Build them around the work instead: the everyday model for everyday tasks, the expensive one only when the job genuinely needs it. And start a new conversation for each new task rather than dragging one thread on forever.
Four Habits That Keep the Bill Under Control
01 · Make the mid-tier model your default. Product descriptions, customer emails, reports, quick questions about a spreadsheet. That’s most of what a retail team does in a day, and a mid-tier model handles all of it. Move up only when it actually falls short.
02 · Start a new chat for each new task. Every message re-reads the whole conversation from the top. A long thread costs more per reply the longer it runs. It feels productive. It’s usually just expensive.
03 · Put your standing documents in a project or workspace. Brand guidelines, tone of voice notes, your product range. Upload them once instead of pasting them into every new chat. It won’t slash your bill, but it keeps output consistent and saves you the repetition.
04 · Set a spending cap before you turn on pay-as-you-go credits. Pick a number you’d be comfortable seeing on a statement, and set it in your account settings before credits go live. Never leave it uncapped.
What Changes When You’re Managing Several Seats
Everything above still applies. These four are on the owner.
01 · Decide who actually needs the top-tier model. Reserve it for the one or two people running genuinely long, multi-step agent work. Give everyone else the standard seat. Buying the premium tier across the board because it sounds like the better option is exactly how budgets disappear.
02 · Set the organisation spend cap before you switch on credits. Most platforms let an owner enable pay-as-you-go access for specific people or the whole team, set a monthly ceiling, and have it cut off automatically once that’s hit. Do that first, then enable access, not the other way round.
03 · Write a one-line model rule and put it in onboarding. Mine looks like this:
Mid-tier model by default. Move up for genuine analysis work. Reserve the top-tier model for named workflows only.
Most overspend isn’t recklessness. It’s nobody ever having said this out loud.
04 · Check usage once a month. Most platforms give owners a usage dashboard. If someone’s consistently at their cap, find out why before increasing it. Sometimes it’s a model habit that needs fixing. Sometimes the work genuinely warrants a bigger allowance. Either way, an uncapped account with no oversight isn’t the fix.
Capability Was Never the Constraint
The retailers actually getting value from AI aren’t the ones running everything on the most expensive model available. They’re the ones who worked out which two or three tasks genuinely need it, picked the cheapest model that handles the rest well, and stopped there.
Discipline beats capability, every time.
If nobody on your team has ever agreed on which model does what, that’s worth sorting out before your next billing cycle does it for you. Get in touch and we’ll help you build a model rule that actually holds.