Why Claude's Token Cost Is Higher Than Ever, by Design

The Claude token cost per unit of real work is higher than it has ever been, and the uncomfortable part is that this looks deliberate. The rate card barely moved. The number of tokens it takes to get the same answer went up a lot.

I’ll say the strong version first, and label it honestly as mine. In my own logs, Opus 4.5 finished tasks for roughly a tenth of the tokens a newer default-thinking model spends on the same job. That’s my measurement, not a study. But the public evidence all points the same way, so I’m comfortable standing on it.

The efficiency high point, then the reversal

Opus 4.5, back in November 2025, was the efficient one. Anthropic marketed it around producing the same quality with about 76% fewer output tokens. That was real, and it was good.

Then the direction flipped. The models after it run extended “thinking” on by default, and thinking is billed. Secondary analyses put newer models at meaningfully higher token use for equivalent tasks, one estimate around 30% more, and the framing that stuck with me was blunt: the same rate card can still produce a larger invoice. When the model decides how much to think and thinking is your cost, “same price per token” is a sleight of hand.

The cut nobody announced

Here’s the one that actually bothers me. Around March 2026, the default prompt-cache TTL appears to have quietly dropped from one hour to five minutes. No blog post. No API version bump. A developer reconstructed the change from roughly 119,000 Claude Code API calls and measured about a 26% overpayment that month against the prior baseline.

Caching is where cost optimisation lives. A cache read is billed at a fraction of fresh input, so a shorter TTL means more misses, and every miss bills like new tokens. Trim the TTL and you raise everyone’s bill without touching a single published price. That’s not a bug. That’s a lever.

It keeps going

The pattern has more entries than one post can hold. Developers reported that a tokenizer change on a newer Opus raised token consumption 32 to 47% for identical work. Third-party frameworks that routed around API pricing got blocked. And in September 2026 Anthropic announced a “25% increase” to weekly limits that, measured against the temporary boost it replaced, worked out to roughly a 17% cut. A cut dressed as a gift.

Any one of these you could wave off. Stack them, and there’s a clear direction: more tokens per task, more misses per session, more friction for anyone trying to spend less. All while the headline price stays quotable.

Why I think it’s structural

Anthropic is burning at a scale that only closes if revenue per user climbs. Thinking-on-by-default climbs it. A shorter cache TTL climbs it. A pricier tokenizer climbs it. None of these require a price hike you’d notice on the pricing page, which is exactly why they work.

I still use Claude. It’s very good. But I’ve stopped assuming the token count is a neutral fact about the model, and started treating it as a number someone at the company has an incentive to grow. If you’re paying per token, you should too.