A Division Problem
On a Saturday morning, Claude Code subscribers received an email from Anthropic: starting September 14, standard weekly limits will be "permanently increased by 25%."
It sounds like a gift. But inside the wrapping paper sits a division problem: 1.25 ÷ 1.5 ≈ 0.833. For the past three and a half months, weekly limits have been on a +50% promotional tier (started May 13). The promo expires August 31, so relative to "today," the new standard is a net −17%. The official announcement did the math transparently, then added a line that reads almost like performance art:
> "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code. We're working on exciting changes that will make it feel like you're getting more from Claude."
*Feel like.* As in, it will feel like more.
The Hacker News reaction was predictable. The thread "Claude Code reduces it's weekly limit by 17% – compared to today" gathered 64 points and 46 comments in two days, with commenters audibly doing arithmetic. One wrote: "This is very sneaky wording, hoping most people won't notice the difference between getting more and *feel* like you're getting more."
Timeline
- May 13 — +50% promotion begins (originally scheduled to end July 13)
- Mid-July — first extension, pushed to August 19
- August 18 — second extension, pushed to August 31
- August 29 — announcement: permanent +25% from September 14
- August 31 — promotion ends
- September 14 — new standard takes effect: net −17% vs the promo period
- Promotion terms: https://support.claude.com/en/articles/15910845-claude-code-may-august-2026-weekly-limits-promotion
- @ClaudeDevs tweet thread (via HN): https://news.ycombinator.com/item?id=49491795
- HN limits discussion: https://news.ycombinator.com/item?id=49506519
- GitHub issue #66504: https://github.com/anthropics/claude-code/issues/66504
- HN session link discussion: https://news.ycombinator.com/item?id=49498201
- CNET lawsuit coverage: https://www.cnet.com/tech/services-and-software/anthropic-sued-alleged-false-advertising-claude-max-subscription-usage-limits/
Why the Cut
The official rationale, compressed into one sentence: "thanks for helping us figure out what we can sustainably serve in the future." Translated: demand is surging, compute is expensive, and the promo is over.
Community readings were blunter. One commenter: "everyone's been running on the promo, demand went to the moon, the music has to stop eventually." Others speculated about tightening spending before an IPO; some quipped that even SpaceX is quoting prices for Cursor.
The genuinely stinging detail: Anthropic has never published absolute weekly limit numbers. Pro is $17–20/month, Max 5x is $100/month, Max 20x is $200/month — but the official 5x and 20x definitions apply only to 5-hour session windows, leaving the weekly limit an unknowable black box. In June, a class-action lawsuit (Northern District of California) targeted exactly this opacity: plaintiffs claim Max 20x actually delivers only "six to eight times" Pro's limits, not twenty, describing the platform as "a black box." User experiences diverge wildly — one $200/month subscriber burns through it "in two days," while another on the same tier runs "all day, never hitting a wall." Same black box, two different lives.
The Second Shoe: Claw Prints in Your Commits
In the same storm, another issue surfaced. Claude Code, by default, appends a claude.ai/code/session_... session link to the bottom of generated commits and PR descriptions. No dialog, no opt-in prompt. On August 30, the related GitHub issue hit the HN front page with 208 points and 229 comments.
The maintainer closed the issue on August 17, stating the links "only appear in web and Remote Control sessions," and offered opt-outs: set "attribution": {"sessionUrl": false} in the config, or the environment variable CLAUDE_CODE_SUPPRESS_SESSION_ATTRIBUTION=1. The trouble: multiple users pushed back on the spot, reporting the link in plain desktop CLI sessions — and since web sessions don't persist a local config file, that opt-out effectively means re-configuring per repository. The closing reply itself was signed "🤖 Generated with Claude Code."
Privacy advocates and pragmatists clashed. Some called it free brand advertising and growth hacking; others noted the link only opens for Anthropic itself ("Sent from my iPhone, but longer"). One pointed question: local sessions are auto-purged after 30 days, yet the link lives forever in public repositories — whose audit trail is being kept by whom?
Why This Matters
Put together, these two episodes mark a turning point for AI coding tools, from "unlimited euphoria" to "metered reality." As coding agents become daily production tools for hundreds of thousands of developers, quota economics and code attribution (watermarking, provenance, copyfraud) become engineering constraints every team must handle. What is the denominator of a subscription price? Where is the boundary of AI-generated code attribution? Two leading companies are voting with their defaults.
Note on the facts: absolute weekly limit values have never been published; multiples here use the pre-promotion baseline as 1.0. The "+25% effective September 14" comes from an August 29 @ClaudeDevs tweet thread (cross-verified via five independent citations; the 16:47 UTC timestamp confirmed by snowflake ID decoding). One clause in the promo terms page implies a fallback to baseline during September 1–13, which conflicts with the tweet's "50% maintained until September 14"; the tweet version is used here and the discrepancy recorded.
Sources