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.
What was checked, and what could not be.
What I verified directly against Anthropic's own documentation (platform.claude.com, fetched Sept. 23, 2026): the exact Opus 5.5 pricing figures ($4/$20 per million input/output tokens; cache reads $0.20/M; 5-minute cache writes $5/M; 1-hour cache writes $8/M; batch processing $2/$10); that Opus 5's prior rate was $5/$25; that the default effort level changed from high to medium; the direct quoted language that Opus 5.5 "tends to think more per turn than Claude Opus 5" at a fixed effort setting; the mechanics of thinking blocks being tied to the producing model, dropped (and unbilled) when a target model can't read them, and the existence of a documented server-side fallback/retry system for declined requests; and the four listed breaking changes bundled into the release (thinking cannot be disabled, forced tool_choice is rejected, thinking blocks are tied to model and conversation, and the older computer_20251124 tool is unsupported on the API and Google Cloud).
What I could not verify directly against Anthropic's own documentation, and attributed to secondary reporting instead: which specific fallback model each safety category routes to (cybersecurity to Opus 4.8; biology and frontier-LLM-development flags to Opus 5) comes from The New Stack's Sept. 22, 2026 coverage, corroborated by a summary of that same reporting from pasqualepillitteri.it. I did not independently confirm Opus 4.8's current per-token price, so the claim that a cybersecurity-flagged reroute costs more than Opus 5.5's rate is not verified here; the claim that a biology/frontier-flagged reroute to Opus 5 costs more is verified, because Opus 5's $5/$25 rate is confirmed on Anthropic's own pricing page and is higher than Opus 5.5's $4/$20.
I was not able to retrieve full article text for two pieces referenced in earlier research (The New Stack's prompt-caching piece and its "four things your agent depends on" piece, both bylined Sept. 23, 2026); repeated direct fetches returned only headline, author, and date. I relied on a separate web search that returned a sourced summary of that same reporting rather than inventing quotes I could not retrieve, and I have not treated anything from that summary as a direct quotation.
The "closer to 40%" and "30% faster output" figures are Anthropic's own claims about its own model, as reported by Benzinga and VentureBeat (both Sept. 22, 2026). I found no independent, third-party benchmark in this reporting cycle measuring the token-reduction claim directly, so it is treated here as an unaudited vendor estimate rather than a verified figure. The Digital Applied breakdown (29% for cache-heavy workloads, roughly 20% otherwise) is that outlet's own analysis, not Anthropic's, and is labeled as such.
The IPO valuation (roughly $2 trillion) and the November timing come from Wall Street Journal reporting as relayed by PYMNTS and Cryptobriefing (Sept. 18-21, 2026); I did not have direct access to the original WSJ article behind its paywall and am citing the secondary reporting rather than the primary piece. The same applies to Anthropic's reported revenue run-rate figures ($9B end of 2025 to $65B by end of July 2026).
The Harvey/Kimi K3 example is used only as industry context for token-cost pressure generally, not as evidence about Anthropic specifically; it is sourced to Dataconomy (Aug. 21, 2026) and TechFundingNews, and the margin figures (50% to -50%) are Harvey's own disclosed numbers, not independently audited.
I did not find any evidence, in Anthropic's release notes or in the secondary reporting reviewed, for two of the specific mechanisms proposed at the outset of this research: a change to context-cache expiry windows, or a change to tokenizer efficiency. Neither is mentioned anywhere in the sourcing gathered for this piece, so neither appears in the article above.