Anthropic shipped Claude Opus 4.7 on Wednesday. The sticker price didn't change. It's still $5 per million input tokens and $25 per million output tokens, same as Opus 4.6. If you glance at the pricing page and move on, you'd assume nothing happened to your bill.
The thing that actually happened is that the new tokenizer produces up to 35% more tokens for the same input text. Same pricing card, more tokens per request, your actual bill goes up. I've read it called a "tokenizer tax." That's a polite phrase for it.
I'm not mad about this, exactly. It's a real model upgrade with real capability gains. What I'm mad about is the way the pricing was communicated, and what that tells me about the pricing model Opus 4.7 was designed to live inside.
In a normal SaaS tool, "sticker price" means something. Your Dropbox bill is $12 or it isn't. Your Notion workspace is $10 per seat or it isn't. The vendor can raise the number, and when they do, you see it, and you decide if it's still worth it.
AI inference isn't like that. The real price has three knobs, and the vendor controls all three.
The vendor can hold knob 1 flat and turn knobs 2 and 3, and your bill rises without anything appearing to have changed. This is not a hypothetical. This is what happened last week. A tokenizer swap that nobody announced loudly increases the effective cost of Opus 4.7 by up to 35% for the same input, and the headlines call the pricing "unchanged."
If you're running a business on top of this, that's a problem. Not because Anthropic is doing anything dishonest (they shipped the numbers, they published the API pricing, the information was there for anyone who looked). The problem is that the structure of the pricing lets the effective cost move without the communicated cost moving. You cannot budget against a number that doesn't capture the variable.
You'd think a flat subscription would insulate you from this. Pay $20 a month for ChatGPT Plus, pay $20 a month for Claude Pro, and the vendor absorbs the token-count changes. Problem solved, right?
Not really. It moves the problem from dollars to rate limits. Here's how it actually plays out: when a new model emits more tokens per request, the subscription vendor has two choices. Option A: eat the cost increase and watch their margin shrink. Option B: tighten the rate limits on the subscription so you hit "you've used your limit for today" sooner. Every vendor eventually picks Option B. I've watched it happen on ChatGPT Plus at least three times in the last 18 months. Same $20 a month, less usage allowed inside it.
The subscription hides the tax. It does not remove it.
And then there's the tier-hopping move. Notion bundled AI into Business-only in 2025, which meant anyone who wanted Notion AI on the Plus plan at $10 a seat suddenly had to upgrade to Business at $20 a seat. For a 5-person team that's a jump from $50 a month to $100 a month for the same core features. The AI "price" didn't go up on paper. The tier you were forced to be on did.
BYOK (bring your own key) means you have your own account with Anthropic or OpenAI or Google, and you paste your API key into whatever tool you're using. The tool makes requests against your key, and the bill from the AI provider comes directly to you.
People usually argue for BYOK on privacy grounds, and that argument is real. The tool vendor can't see your prompts because they never touch them. I wrote a whole essay on that angle. But the pricing argument is just as strong.
When you're on BYOK, the tokenizer tax shows up in your invoice. Your Anthropic bill at the end of the month says "you used 12.3M input tokens and 4.1M output tokens, here's the line item." If last month's same work used 9M input tokens and the tokenizer changed under you, you see the delta in plain text. You can measure it. You can switch models if the new tokenizer is bad for your workflow. You can complain on Twitter with actual numbers.
On a subscription, none of that is visible to you. You just notice you hit the rate limit faster.
BYOK doesn't eliminate pricing changes. It just makes them impossible to hide. Every cost increase shows up on a line item you're looking at, in a currency you're paying.
If you're a salaried employee whose company pays for Claude Team, none of this matters to you. Your employer eats the tax. Fine.
If you're an indie founder paying for your own tools, this matters a lot. I spent 2025 keeping a spreadsheet of my AI spend, and the thing I kept running into was that I couldn't forecast it on subscription tools. I'd pay $20 for ChatGPT Plus in January, hit the rate limit in February because they changed the model under me, upgrade to Plus Pro at $200, realize I didn't need Plus Pro, downgrade, start the cycle over in April.
When I switched my strategic-planning work to Projelli (the local-first AI workspace I built, BYOK to Anthropic directly) my actual AI spend for that workflow dropped to about $6 a month. Not because the AI got cheaper. Because I was now paying for exactly what I used, which turned out to be less than the flat fee assumed.
The other direction is also true. If you use AI heavily, a subscription can be a better deal than BYOK for a while. Then the vendor tightens the rate limits and it isn't. BYOK is stable in a way subscriptions aren't. Stable meaning: the incentive alignment between you and the provider doesn't flip overnight because the provider changed a tokenizer.
I'm a tool vendor. Projelli sells for $49 one-time, and I don't charge anything on top for AI usage because I never touch the AI request. You bring your own key, your key pays your provider, I'm not in the request path. That's a deliberate decision and it cost me revenue. I could easily have built Projelli with a $15 a month "AI credits" layer that marked up the inference by 40%. Every other workspace tool does. I chose not to, because I think the founder-vendor contract around AI pricing should be honest.
Here's what I think a tool vendor owes you on AI pricing, in order of importance:
Very few tools do 4 out of 5 of these. I can count on one hand the ones that do all 5. (Cursor does most of them. Obsidian community plugins do all 5, because they're community plugins and have nothing to sell you.)
Three concrete moves if you're a founder and you care about not getting tokenizer-taxed into confusion:
Open an API account with at least one provider directly. Takes 5 minutes at console.anthropic.com or platform.openai.com. Load $10. Now you have a billing relationship with the primary source, not a reseller.
Audit your AI subscriptions. How much are you paying across ChatGPT Plus, Notion AI, GitHub Copilot, Cursor, whatever else? For how much actual usage? If the aggregate is over $80 a month, BYOK is probably cheaper for you even if your usage is heavy.
Pick one workflow to move to BYOK as a test. Not everything at once. One workflow. See how the bill looks after a month. If the math works, move the next one.
This isn't about the $6 a month I'm saving on Projelli versus Notion AI. It's about owning the math. The specific thing that made me start Projelli was noticing that every time a vendor changed a tokenizer or a tier or a rate limit, I was buying a product I hadn't agreed to buy. BYOK is the only pricing model where that can't happen to you, because the provider is billing you directly, and the line item is in your own account, in dollars you chose to spend.
The tokenizer tax doesn't go away. The visibility does.
See Projelli, local-first AI workspace for indie foundersJameson Daines builds Projelli on weekends and evenings around a Senior Product Designer day job. Read the local-first argument or get Projelli at projelli.com.