April 2026 had three Claude announcements, and only one of them actually moved your plan limits

Two. Claude Mythos Preview launched April 7, 2026 as a research preview limited to Project Glasswing partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks). It is not available on Pro or Max plans. Claude Opus 4.7 launched April 16, 2026 as the generally available top-tier model, replacing Opus 4.6 across claude.ai, Claude Code, the API, Bedrock, Vertex AI, and Microsoft Foundry. Same $5/$25 per million input/output tokens as Opus 4.6. A third event on April 23 was not a launch but a postmortem: Anthropic published the Claude Code quality postmortem and reset usage limits for all subscribers.

Anthropic's Opus 4.7 launch post on anthropic.com/news/claude-opus-4-7 does not mention rate limits at all. The 'What's new in Claude Opus 4.7' developer doc only addresses tokenizer changes and breaking changes to sampling parameters and extended-thinking budgets. There is no public statement that subscriber denominators changed at the moment 4.7 went GA. The fields you read on /api/organizations/{org}/usage are the same as on April 15 (five_hour, seven_day, seven_day_sonnet, seven_day_opus, seven_day_oauth_apps, seven_day_omelette, seven_day_cowork, plus extra_usage). Only the fill rate changed.

Three product-layer changes between March 4 and April 16 had degraded Claude Code's quality. The April 23 postmortem at anthropic.com/engineering/april-23-postmortem confirmed: a March 4 default-effort change from high to medium (reverted April 7), a March 26 prompt-caching bug that cleared thinking history every turn instead of once (fixed April 10), and an April 16 verbosity instruction that hurt coding quality (reverted April 20). The verbosity test alone produced a 3% quality drop on both Opus 4.6 and 4.7. Resetting subscriber quota was the make-good. The literal text from the postmortem: 'As of April 23, we're resetting usage limits for all subscribers.'

Every Window struct's utilization field returned to zero on the server side, regardless of how high it had been the moment before. The shape did not change: still seven Option<Window> fields plus an extra_usage block. The resets_at timestamp on each window also rolled forward to a fresh 5-hour or 7-day window starting at the reset moment. ClaudeMeter accounts polling the endpoint that day saw five_hour, seven_day, seven_day_opus, and seven_day_sonnet drop simultaneously to a fraction near 0.0 and watched their resets_at timestamps jump forward by hours or days.

No. Mythos is research-preview-only via Project Glasswing and is not on the Pro or Max consumer plans. There is no field for it in /api/organizations/{org}/usage. If your account does not have Glasswing access, Mythos traffic cannot exist on your account. Even partners with access pay $25 input / $125 output per million tokens billed under the $100M credits Anthropic committed to Glasswing, not under their Pro or Max subscription. The April 7 announcement does not change a single number you see on claude.ai/settings/usage.

Two compounding causes from the Opus 4.7 release notes (platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7). One: the new tokenizer maps the same English text to 'roughly 1x to 1.35x as many tokens' compared to 4.6 (up to about 35% more, varying by content). Server-side accounting uses the new tokenizer's count, so the same prompt fills seven_day_opus faster. Two: 4.7 added an xhigh effort tier that did not exist on 4.6, and adaptive thinking is the only thinking-on mode. Adaptive thinking generated tokens count as output tokens. Same prompt, same agentic loop, more tokens written to the Opus-only weekly float. There is no per-version field, so you cannot back out the 4.7 vs 4.6 share after the fact.

No. The UsageResponse struct in claude-meter/src/models.rs lines 18-28 declares seven Option<Window> fields plus an ExtraUsage. None of them name a model version. Opus 4.7 traffic, Opus 4.6 traffic, Sonnet 4.6 traffic, and any future model launched on the consumer plan all serialize into the same fields. The extension at extension/popup.js lines 60-63 already optional-chains the model-specific weekly fields, so accounts that never had a seven_day_opus row simply do not render one. The Apr 23 reset also required no code change: utilization went from 0.91 to 0.02 on the same field.

Anthropic shipped Claude Code v2.1.116 on April 20 with the verbosity instruction reverted, and the postmortem (published April 23) lists all three quality issues as resolved. The reasoning effort default is now xhigh on Opus 4.7 and high on every other model. The thinking-history clear bug was fixed April 10. Whether your specific session feels back to normal depends on which of the three changes was hurting your workload most. None of these fixes change which fields gate your account's rate limit.

The resets_at field on the seven_day_opus Window struct is the absolute UTC timestamp. It is returned as a Option<chrono::DateTime<chrono::Utc>> in claude-meter/src/models.rs line 6. After the Apr 23 reset, the timestamp on every account's seven_day_* fields jumped forward to seven days from the reset moment. ClaudeMeter prints it relative ('resets in 6d 4h') in the menu bar; the raw ISO timestamp is on the localhost bridge at 127.0.0.1:63762/snapshots if you want to script against it.

If you only check the bar once a day, the Settings page is fine. The reasons people install a tracker after April 2026 are: (1) the 4.7 tokenizer changes how fast the Opus float climbs and you want to recalibrate your sense of 'how much weekly is left', (2) the Apr 23 reset moved everyone's resets_at forward, so the old wall-clock heuristic ('I started Tuesday, I'm safe through Tuesday') no longer matches the server, and (3) Settings does not warn you as you approach a limit; ClaudeMeter polls every 60 seconds and the menu bar pip turns red at 80%.