Anthropic Called It a 25 Percent Increase. Do the Math and It Is a 17 Percent Cut.

On August 29, Anthropic's developer account announced that starting September 14, standard weekly usage limits in Claude Code would rise 25 percent for Pro, Max, Team, and seat-based Enterprise plans. That same day, the company posted a follow-up: compared to what users have right now, the change is actually a 17 percent reduction. Both numbers are correct. They are just measured from different starting lines, and Anthropic led with the flattering one.

Here is the arithmetic. Index the pre-promotion standard weekly limit at 100. Since May 13, Claude Code has run a temporary 50 percent boost, putting current usage at 150. On September 14 that boost ends and a new, permanent standard limit lands at 125, which is 25 percent above the original 100 baseline. It is also 125 divided by 150, or about 83 percent of what users have today, which is where the 17 percent figure comes from. Digital Applied's breakdown of the announcement lays out the same indexed math and adds a detail that matters: as of the morning after the announcement, Anthropic's own support documentation still said the boost expired August 31, with no mention of the September 14 date or the permanent 125 figure at all. Two arms of the same company were telling customers different things about their own usage caps.

Developer and blogger Manton Reece put it plainly after reading the announcement: "I had to read it a few times. It's actually bad news," adding that Anthropic communicated the change "in the most convoluted way possible." On Hacker News, commenters worked through the same 150-to-125 math within hours, according to Digital Applied's account of the thread, and some initial coverage muddled the two percentages entirely, treating "50 minus 25" as if that were the real cut instead of the actual ratio.

The steelman for Anthropic is worth stating plainly, because it holds up better than the pile-on suggests. A temporary promotion is, by definition, temporary. Software vendors run trial pricing and usage boosts that expire constantly, and nobody calls a gym membership a "cut" when the introductory rate ends and the standard rate begins. Here the standard rate itself went up, permanently, by a real 25 percent over the original baseline, which is a genuinely better deal than reverting to 100 the way the boost's own terms always implied it eventually would. Leading a company announcement with the number that reflects the permanent, structural change rather than the expiring promotional one is not obviously dishonest. And to Anthropic's credit, the company corrected its own framing the same day, in public, before this became a slow-drip trust problem.

But the correction had to happen at all, and that is the part worth sitting with. XenoSpectrum's reporting notes the documentation mismatch persisted with no stated cutover time or rollout detail even after the clarifying post went up. A company selling infrastructure that engineering teams build real, revenue-generating workflows on is also selling the reliability of its own numbers, and a customer base sharp enough to catch a ratio error in an afternoon is also sharp enough to remember it the next time a pricing page changes.

The part I would flag for any business running meaningful volume through Claude Code, or evaluating whether to: never take a vendor's headline percentage as your planning input, on limits or on price. Help Net Security's coverage of the boost's earlier extensions shows this 50 percent bump has already been extended multiple times since May, each time reframed slightly differently. Track your own number instead: how many real tasks your team completes per week under the limits you actually have, measured before September 14 and again after. If that number drops meaningfully once the boost lapses, that is your actual signal to start routing routine work to a cheaper or local model and saving the premium agent loops for what needs them, not the vendor's press release.

And notice what forced the correction. No regulator required Anthropic to disclose the math in plain terms. A blogger and a Hacker News thread did the arithmetic, said so publicly, and the company fixed its own framing within a day because the alternative was worse for its reputation. That is market accountability working faster than a mandated disclosure rule ever would, and it is worth crediting even when the company that got caught is one I use myself. If you want help sizing what your team actually needs from Claude, GPT, or an open model instead of guessing from a vendor's percentage, that is a conversation Mojo AI Services has for free.