Opus 5.5 Is 20% Cheaper. The Other 20% Is Anthropic Asking You to Let It Think Less.
Anthropic's flagship price cut is real and checks out against its own price list. The "closer to 40%" savings number the company is showcasing three weeks before a roughly $2 trillion IPO roadshow depends on assumptions its own documentation quietly contradicts.
Am I The Idiot — the strongest case against our own take.
The strongest case against this piece: everything in the "hidden cost" argument here can also be read as normal software evolution, not a trick.
Every model release changes behavior somewhat. Anthropic's own documentation is unusually candid, on its own public docs page, about the model thinking more per turn at a fixed effort setting; a company trying to obscure that would not publish it under its own heading with a link to a dedicated mitigation guide. Anthropic was not obligated to disclose that at all, and doing so cuts against the idea that the 40% figure is engineered to mislead.
The default-effort change from high to medium is also plausibly a genuine capability improvement rather than a lever. If the model is simply better trained, needing less deliberation to reach the same answer quality is what "better" is supposed to look like, and multiple outlets (VentureBeat, TechCrunch, both Sept. 22, 2026) reported Opus 5.5 matching the pricier Fable 5.1 tier on agentic benchmarks. That is a specific, testable claim, not just marketing copy.
The safety classifier reroute is a legitimate safety feature, not an obvious monetization tool. Anthropic gains nothing evident by routing cybersecurity- or biology-flagged requests to different models; those categories are a small fraction of ordinary traffic, and Anthropic bears real engineering cost maintaining classifier coverage and fallback compatibility across model versions, work it did not have to take on.
The competitive framing cuts against the cynical reading too, not for it. If Anthropic could safely charge more and still tell its growth story, a genuine price war with OpenAI, which cut its own prices roughly in half the same day, and with xAI, already pricing Grok 4.7 at $2 input, is exactly the kind of external pressure that forces real cuts whether a company wants to make them or not. A firm preparing to sell public-market investors on a growth narrative generally wants to protect margin, not compress it into a three-way price war three weeks before a roadshow. That Anthropic cut anyway reads at least as plausibly as competitive necessity as it does as manufactured generosity.
The honest weak point in this piece is that "the marketing number rounds up from the verifiable one" is a fairly ordinary complaint, not evidence of a coordinated strategy to raise effective prices while claiming to lower them. Anthropic's unconditional cuts, 20% flat on the base rate and 60% on cache reads, are real, larger than what some competitors have offered on a flagship model this year, and require no asterisk at all. If the underlying goal were to extract more revenue while claiming otherwise, cutting the single number that a developer can verify against a price list in five seconds would be a strange way to go about it.