On Tuesday, Anthropic put Claude Opus 5.5 on its pricing page at $4 per million input tokens and $20 per million output — 20 percent below Opus 5 — and then did the more interesting thing three rows down. Cache reads fell from $0.50 to $0.20 per million tokens. That is a 60 percent cut, three times deeper than the headline number. Cache writes, by contrast, went from $6.25 to $5.00. A 20 percent trim.

The trade press read the top line. The strategy is in the asymmetry.

The Lever They Pulled Hardest Was the Price of Remembering

Fresh inference is thinking. A cache read is remembering — the model re-reading context you already sent and already paid to deposit. And cached context is not some incidental artifact of your integration. It is the system prompt with your compliance rules in it. It is the retrieval index of your policy manual. It is the 200,000-token agent transcript your coding tool has been accumulating since March. It is the vocabulary your organization uses and no other organization does.

Anthropic just made remembering three times cheaper than it made thinking. Depositing the corpus still costs you — $5.00 for a five-minute cache write, $8.00 for an hour — but once it is in, the meter runs at a fifth of a dollar per million tokens.

“We cached about 300 million tokens of claims history and policy language,” a platform engineer who runs the model gateway at a mid-sized insurer wrote in a Slack DM the afternoon the pricing page updated. “That’s not a workload. That’s a filing cabinet. Tell me the price of moving it.”

She could find out. It would be the price of writing it all again somewhere else, at the un-discounted cache-write rate, plus the weeks of re-tuning that go with it. The discount is real. That is precisely what makes it a moat.

The Bill Per Finished Task Barely Moved

Give Anthropic credit for the honesty of its own claims: more than 30 percent faster, and about 40 percent cheaper on typical workloads. Note the qualifier. The 40 percent is a claim about default effort and about the model spending fewer tokens per task, not about the rate card. On the rate card, output is down 20 percent.

Latent.Space’s tracker put Opus 5.5 at $5.98 per task on the Intelligence Index at maximum effort, against $5.86 for Opus 5. A hair higher. Rate down, consumption up, and the two roughly canceled out.

Which means the number every procurement deck is denominated in — dollars per million tokens — has become a vanity metric. The number that matters is dollars per finished thing: per closed ticket, per reviewed contract, per merged pull request, per claim adjudicated. Anthropic cut the first and quietly let the second float. Anyone who circulated the 20 percent this week as a cost reduction for buyers owes their CFO a correction, not a victory lap.

Nobody to Sue, Which Is the Interesting Part

There is no merger here. No exclusive-dealing clause, no captured standards body, no consent decree to appeal. The lock-in was built out of a price sheet, and it is legal in every jurisdiction that matters.

Competitors matched the headline. Claude Fable 5.1 sits at $10 in and $50 out; GPT-6 Astra at $10 and $50; Opus 5.5’s own fast mode at $8 and $40. What none of them appear to have matched is the cache policy — the lever that decides whether a customer’s accumulated context is an asset they own or a deposit they hold.

For the antitrust-first crowd, that is an uncomfortable fact: there is no defendant to name. For the markets-fix-everything crowd, it is worse. A buyer holding 300 million tokens of cached context and a re-tuning bill does not have a credible exit, and a market where exit is irrational is a market where the winner stops having to try. Watch the next launch to see what Anthropic declines to discount.

The vendors selling model-routing layers — “swap frontier models in an afternoon” — have the most immediate problem. The swap works fine right up until you price the cache. Then portability turns out to be a slide, not an architecture. And the analyst who modeled falling inference prices as a straight rebate from labs to enterprises should mark the position to market: a slice of that rebate is being paid back as captivity.

None of this requires villainy. It is what any competent business does when the product is commoditizing and the customer’s data is not. Anthropic found the one line item where a discount buys more than it costs, and priced it accordingly.

The bank hands you a toaster. The lab hands you twenty-cent cache reads. Both are counting on you to keep the account open.

Sources