Anthropic released a new AI model yesterday and priced it 60% cheaper than their flagship. I saw that headline and had the same reaction most people probably did: that’s a deal. Then I read the fine print, and it’s a much more interesting story than a price cut.
Buried in the release notes is a detail that changes what “60% cheaper” actually means. And the mechanic behind it isn’t unique to Anthropic — it’s sitting in the pricing page of pretty much every AI vendor you or your clients are evaluating right now.
What actually happened
The new model is called Sonnet 5. It launched at $2 per million input tokens and $10 per million output tokens through the end of August, moving to $3 and $15 after that. Compare that to Anthropic’s flagship model, Opus, which runs $5 and $25 — so on paper, Sonnet 5 is 60% cheaper.
Here’s the part that doesn’t make the headline: Sonnet 5 shipped with a new tokenizer — the rulebook that chops text into the “tokens” vendors bill you for. According to Anthropic’s own release, that new tokenizer can produce up to 1.35 times more tokens for the exact same piece of text compared to the old one.
Sit with that for a second. If the same email, the same memo, the same document now gets chopped into 35% more tokens than before — and you’re billed per token — a chunk of that advertised discount just evaporated. Anthropic has actually been upfront about this: they’ve said the introductory price was set specifically to keep the switch roughly cost-neutral because of the tokenizer change. Translation: the sticker price dropped, but the unit you’re being charged for got smaller, so the real cost to get a task done didn’t fall anywhere near 60%.
This is the same question as APR on a loan
Two lenders can both advertise a “5% rate,” but if one compounds daily and the other compounds monthly, those aren’t the same 5%. You’d never sign a loan off the headline number alone — you’d ask how it compounds. This is the identical question, just wearing an AI costume: is this the stated rate, or the effective rate?
I want to be clear that I don’t think Anthropic did anything shady here — they disclosed the tokenizer change and priced around it. But that’s exactly the point: even when a vendor is being transparent, the headline number still isn’t the number you should budget against. And not every vendor is going to be this upfront about what changed under the hood.
What this means for you
Back in my PwC days, I sat through more vendor contract reviews than I can count, and the pattern was always the same: someone brings a proposal into the room with a lower rate than the incumbent, everyone gets excited, and the actual due diligence — what unit is this rate measuring, and has that unit changed — happens two meetings too late, if at all.
Two things worth doing with this. First, any time you’re comparing AI vendor pricing — for a client’s tech stack or your own firm’s tools — don’t compare rate cards. Get a cost-per-task number instead: run the same representative document or workflow through both options and compare what it actually costs to get the job done. Second, if your firm tracks AI spend internally, treat any vendor’s model or pricing change as a mandatory re-baseline trigger, not a footnote. The same dollar figure in your tracking sheet can now represent a meaningfully different amount of real usage.
Key Takeaways
- Sonnet 5 launched at $2/$10 per million tokens — a genuinely lower advertised rate than Anthropic’s flagship model and its direct competitors.
- A new tokenizer means the same text can cost up to 35% more tokens than before, so real savings are smaller than the 60% headline suggests.
- Any time you see a vendor’s rate card, ask the APR question: stated rate, or effective rate? Get a cost-per-task comparison before drawing a conclusion.
- If you track AI spend internally, treat every vendor pricing or model change as a required re-baseline — not a footnote.
The practical test for your next vendor conversation: don’t ask what the price per unit is. Ask what it would cost to run last quarter’s actual work through the system. That’s the number that tells you the truth — the rate card never will.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Price Cut That Isn’t
Leave a Reply