[Bug] Misleading "Resets at HH:MM" label describes sliding window as resetThis is feedback on the /usage screen's "Resets at HH:MM" label, but the broader issue is product trust.

I am a Korean developer, paid subscriber for over a year, currently
extra-usage 120%. I am writing because Anthropic's credibility is leaking
faster than the company appears to realize.

---

## The immediate bug

The /usage screen labels the 5-hour Session window as "Resets at HH:MM."
The actual behavior is a sliding window — at the displayed time only the
oldest tokens expire, leaving 40 to 90 percent of usage in place. "Reset"
universally means the counter returns to 0. Using the word for behavior
that never returns to 0 is not a wording mistake. It is a misleading
label that exploits the universal meaning of the term.

Industry baseline:

• GitHub's X-RateLimit-Reset is honest because their window is tumbling.
• Stripe and AWS describe their rolling windows as "rolling."
• Anthropic is the only major provider using "Reset" for sliding behavior.

The deception has stood for over a year despite repeated complaints
(anthropics/claude-code#51222 remains open). The absence of a real-time
weekly-cap dashboard compounds the problem — users cannot verify their
own consumption, only react to surprise.

---

## This is not isolated. It is part of a pattern.

In the past 3 weeks alone, Anthropic shipped:

• Opus 4.5 / 4.6 removed from Pro+. Opus entirely removed from Pro.
• Opus 4.7 launched with tokenizer change — list price unchanged, real

cost up. Same family of deception as the "Reset" label.

• 7.5× multiplier framed as "promotional," reverting to 15× from June.
• Copilot Opus 4.7 capped at 192K context, out of the model's 1M

actual capacity. Premium users charged premium prices for 20% of
the product.

• GitHub paused new Copilot signups on April 20 citing Anthropic

compute cost spikes.

GitHub Discussions #189999, #192963, #193237 are all live, all loud.

The "Reset" label is not a forgotten UX detail. It is consistent with
a 3-week pattern of opacity around what users actually receive: prices
up, capability down, no advance notice, no honest labeling.

---

## The business comparison Anthropic should sit with

Netflix Korea (2016 → today): entered at ₩9,500/month. Invested in
local content first (Kingdom, D.P., Squid Game, The Glory). Korean
content became 30%+ of Netflix's global growth engine. Every price
hike came with visibly more value. Korean user loyalty is now structural.

Disney+ Korea (2021 → today): entered at ₩9,900/month with weak
content, poor subtitles, awkward UX. Raised prices anyway. Korean
subscribers collapsed. Gap with Netflix widened. They are now a case
study in how to lose Korea.

| Axis | Netflix (growth) | Anthropic (now) |
|---|---|---|
| Price | up | up (incl. tokenizer trick) |
| Value | up (originals) | down (Opus removed, context capped) |
| Transparency | advance notice | policy changes shipped without warning |
| Monitoring | clear usage visible to user | no weekly-cap dashboard |
| Labeling | accurate | "Reset" applied to non-resetting behavior |

Anthropic is currently executing the Disney+ Korea playbook, not the
Netflix one.

---

## Why Korea specifically — the asymmetric exposure Anthropic
underestimates

Korea is 1-2% of Anthropic's user base by count. The response pattern to
perceived deception is well-documented and does not look like other
markets:

• Korean consumers do not "cool off." A brand judged dishonest is

permanently dropped. Namyang Dairy (2013) — 10+ years of unrecovered
sales. Oxy Reckitt (2016) — market exit. Japanese consumer goods
(2019 No-Japan) — beer sales -90%, still not recovered. There is no
"win them back with a discount" path.

• Korean IT communities (OKKY, GeekNews, Clien, dev YouTubers) ignite

within days once a credibility issue gets a clean narrative. A
one-word framing — "Reset이 사기" — is exactly the kind that spreads.

• Korean developers have full access to alternatives: ChatGPT (very

high Korean Plus adoption), Gemini, domestic LLMs (Naver HyperCLOVA
X, LG EXAONE), local models via Ollama. Switching cost is near zero.

The asymmetric leverage:

• SK hynix supplies the HBM that NVIDIA H100/H200/B100 require. Samsung

is the second source. Together: effective Korean monopoly on the
memory layer Anthropic's training and serving infrastructure depends
on. Korean public sentiment, supply chain partners, and policy
makers are coupled in a way that does not exist in larger markets.

• This is not a threat. It is a description of how Korean consumer

distrust historically propagates upward into industrial and policy
decisions. Anthropic's structural exposure is much larger than
"1-2% of revenue" suggests.

---

## In hardware that I built

Korean thermal power plant operating systems, reverse-engineered from
ABB, Siemens, Westinghouse, Mitsubishi protocols over 1995–2009. A
label that lies about its function would never ship — the customer
(KEPCO and 5 major utilities) would reject the panel and invoice the
vendor for the lost outage window. SaaS apparently has no equivalent
accountability. That is why this has persisted.

---

## What needs to change

1. Replace "Resets at HH:MM" with language that matches actual behavior:

"Oldest usage expires HH:MM. True 0% only 5h after last call."
"Rolling window — usage decays continuously."
"Earliest tokens expire HH:MM (sliding 5h window)."

1. Ship a real-time weekly-cap dashboard. Its absence is the

single most-cited missing feature among power users.

1. Publicly acknowledge that anthropics/claude-code#51222 is a real

bug and commit to a fix timeline.

1. Stop shipping price increases as tokenizer changes. Stop shipping

capability cuts (Opus removals, context caps) without advance notice.
These are the same class of deception as the "Reset" label, just
wrapped in different surface area.

---

I have been a paying customer for over a year. I am building my own
tools on top of Ollama specifically to reduce dependency on Anthropic
because of incidents like this. I am the canary, not the edge case.

The question is whether Anthropic chooses to fix this now — while it is
one open issue and a few angry power users — or later, when the
Korean-language news cycle ignites and propagates to other markets via
the same supply chain partners Anthropic depends on.

Fix the label. Then fix the pattern.

Additional observation — equity by usage pattern:

A user who closes Claude Code overnight experiences a real 0% the next
morning. A user who runs Claude continuously (my pattern: 14 days
uptime, session essentially never closed) experiences perpetual residual
usage because the sliding window's "last call + 5h" condition is never
satisfied. Same subscription tier, same listed limits, but the heaviest
users — the ones who pay the most attention to and depend most on the
product — receive systematically less. The "Reset" label hides this
asymmetry. Light users see what the label promises. Power users never
do. This is not just a labeling problem; it is a pricing fairness
problem disguised as a labeling problem.