This Time the Two Kinds of "Increase" Were Written Separately, and Last Time They Were Not
On September 17 this site covered a case where Anthropic's announcement emphasized a "permanent 25% increase" to Claude Code's weekly limits while letting a temporary 50% bonus expire at the same time — the two together left users with roughly 17% less than the week before. A developer on X added a community note to the original post, and the company deleted and reposted it, acknowledging the decline. This time the wording is much cleaner, and the contrast is worth recording: **the five-hour limit rose 20%, which means the allowance itself got bigger; while "five-hour and weekly limits go roughly 25% further" means the same allowance now buys more work.** Those are two different things, and Anthropic wrote them separately this time. Why insist on the distinction? Because the second one depends on your workload. **If only the 25% gets reported, readers will believe the allowance rose 25%, when it actually rose 20%** — the remainder has to be earned back by the model being cheaper, and how much of it you get varies by user.
"40% Less" Is a Composite Claim; the Checkable Part Is 20%
The price list is hard: $5/$25 to $4/$20, or 20%. Cache reads from $0.50 to $0.20, or 60%. The 40% is composite — it combines the lower unit price, fewer tokens per task, and less compute to serve. The last two are not on the price list and depend on what your tasks look like. So the prudent approach for procurement or budgeting is: **assume 20% conservatively, then measure the composite figure on your own tasks.** There is, though, a way to predict which end you will land on ahead of time: **the size of the cut depends on your cache hit rate.** Anthropic itself says cache reads make up the majority of cost in agentic and coding workloads, and cache reads fell 60%. So long sessions with large contexts read repeatedly will land clearly toward the 40% end, while short, uncached one-off calls get 20%. That ratio is something you can look up in your own billing. A judgment from this site's July 13 piece on an earlier round of token price cuts still holds: **being able to save money is not the same as saving it** — cache hits, model tiering and usage attribution are what actually bring a bill down. Today's cut amplifies that, because the cut on the cache tier is three times the cut on the headline price.
What Engineering Teams Should Notice Is Not the Price Cut, It Is Two Changed Defaults
First: **the effort setting now defaults to medium instead of high.** Anthropic says medium matched or beat Opus 5 at high on coding and knowledge-work evaluations, and that may well be true. But the implication is that **after upgrading, the same call runs on a different reasoning budget by default.** Combined with adaptive thinking being impossible to disable, this is a change that requires deliberate regression testing for any pipeline that needs stable, reproducible output. It is not a transparent upgrade. Second, and harder: **safety classifiers can reroute requests to older models mid-workflow**, particularly on cybersecurity tasks. Which means that while you believe the whole chain is running on 5.5, some requests in the middle may not be. For a production system, that means latency and output characteristics can shift without you noticing. Vetted researchers can apply to the Cyber Verification and Life Sciences Verification Programs for broader access, but for an ordinary team the useful step is logging the model identifier that actually responded rather than assuming it equals the one requested.
The Bankable Reset Is a New Mechanism, Not Just More Headroom
Subscribers got more than a higher ceiling this time: they got one rate-limit reset they can spend at a moment of their choosing. That deserves its own note, because it is not a variant of "raise the limit." **It hands part of the allowance-management decision to the user** — you can spend it on a deadline day, during a large refactor, or when you suddenly need to run a batch of evaluations. For bursty usage, one self-timed reset is more useful than a flat 20% increase. The cost is that the responsibility moves with it: whether and when to use it is now your call. Mechanisms like this typically benefit heavy users most, because they know where their peaks are. Two points of wording to close. One outlet reported the five-hour caps as removed while most describe them as raised 20%; this site follows the majority, but the discrepancy is a reminder to trust the limit you actually see in your own account. And fast mode is priced at $8/$40, double the standard rate — **up to 2.5x faster at twice the unit price**, a trade whose worth depends entirely on how expensive your time is. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks. On the same day, OpenAI extended the GPT-6 family downward and cut prices too; this site has a separate piece on that today. Two companies moving on price the same day is not a coincidence.
via: Betanews, The New Stack, Mixed News, Quartz