Claude Code Max usage limits: the published numbers, and the float they actually come from.

No. The help center articles for Pro, Max, and Claude Code repeatedly say Max gives you 'higher usage' or '5x' and '20x' Pro's allowance, but they never name a prompt count, a token cap, or an hours-per-week number. The actual enforcement primitive is a utilization float on claude.ai/api/organizations/{org_uuid}/usage. The float lives between 0.0 and 1.0. When it hits 1.0, the next prompt returns 429. Independent tests (truefoundry, intuitionlabs) converge on ~50 to 200 prompts per 5-hour window on Max 5x and ~200 to 800 on Max 20x before the May 6, 2026 doubling. After the doubling, those become roughly ~100 to 450 and ~400 to 1800. The range is wide because token mix per prompt swings the float by ~3x.

Anthropic announced that the 5-hour rate limit was doubled on May 6, 2026 for Pro, Max, Team, and seat-based Enterprise. Peak-hour throttling was also removed for Pro and Max. The change was tied to a compute deal with SpaceX's Colossus 1 data center. Weekly caps were NOT changed: every announcement and follow-up restated 'the weekly bucket is the same size it was before.' The five_hour bucket on the /usage endpoint now climbs to 1.0 at roughly half the per-prompt rate it did on May 5; seven_day, seven_day_sonnet, seven_day_opus, and seven_day_oauth_apps still climb at the old rate.

Because nobody is publishing a constant. Every blog you read is reverse-engineering by running test sessions and watching when their account 429s. A 'short prompt with a small file edit' burns the five_hour float at one rate. A 'deep refactor with 50 KB of context, adaptive thinking on, Opus 4.7 tokenizer' burns it at three times that rate. Same plan, same week, different prompts, different effective cap. The float is the contract. The prompt counts are downstream estimates that wobble with model choice, context length, and tokenizer version.

Five, exposed as separate fields on the same /usage endpoint: seven_day (all-up), seven_day_sonnet (Sonnet-only), seven_day_opus (Opus-only), seven_day_oauth_apps (Claude Code and other OAuth clients only), and two internal-feature buckets (seven_day_omelette, seven_day_cowork). Plus one dollar-denominated cap on /overage_spend_limit when metered billing is on. Any single bucket at 100 percent fires the 429. The big bar on claude.ai/settings/usage is the seven_day aggregate; the four model-specific and OAuth-specific buckets do not get their own rows in the prose.

Most Max users hit seven_day_oauth_apps mid-week. Claude Code authenticates through the OAuth flow, so every prompt counts toward the OAuth-only weekly bucket on top of every other bucket it touches. Pure-claude.ai-web work skips that bucket entirely. Heavy Opus-on-hard-tasks workloads hit seven_day_opus instead. Mixed teams that share a Max seat hit seven_day_sonnet faster than expected. The five_hour cap is the most visible because it fires several times a day; the weekly caps are the ones that surprise people who think 'Max 20x' means 'I have 20x the headroom forever.'

No. The float is always 0.0 to 1.0 regardless of plan. The plan determines how many tokens push the float by 0.01, not the upper bound of the float. Anthropic uses 5x and 20x as marketing labels for 'approximately 5x more prompts per session' and 'approximately 20x more prompts per session.' Internally there is just one utilization fraction per bucket per account, and the cap behind it varies by plan. ClaudeMeter reads the fraction as a number between 0 and 1, multiplies by 100, and prints '21.0% used' next to a row in the menu bar.

It does not change the plan caps. It adds an eighth gate. Once extra_usage.is_enabled is true on the /usage payload, the next prompt that would have 429'd against a plan cap proceeds at standard API prices and counts against monthly_credit_limit on /overage_spend_limit. When used_credits crosses monthly_credit_limit, out_of_credits flips true and disabled_until appears. So metered billing converts the rolling-window 429 into a dollar-cap 429. Both are visible to ClaudeMeter; the menu bar shows '$ 78.40 / $ 200.00 (39%)' as its own row.

Yes. claude-meter runs as a macOS menu bar app and a browser extension. The Rust core in src/api.rs (lines 8 to 30) calls GET /api/organizations/{org}/usage and GET /api/organizations/{org}/overage_spend_limit once per minute using the cookies the browser already holds. It deserializes into UsageResponse (models.rs lines 18 to 28), and renders each bucket's utilization as a row in the popover. No cookie paste, no token, no telemetry. Read-only against endpoints the Settings page already calls. The CLI version prints the same rows for tmux and Starship status lines.

ccusage reads ~/.claude/projects/<project>/<session>.jsonl on your machine and sums input and output tokens against a model price card. That is a cost calculator for tokens that left your laptop. Plan limits live on Anthropic's servers as the utilization floats above. The two numbers describe different things and were never supposed to match. ccusage is great for 'how much did I spend in API-price equivalents this month.' It cannot read the plan quota because the plan quota is not in the local logs. ClaudeMeter and ccusage answer different questions; running both is reasonable.

Partially. Pro, Max, Team, and seat-based Enterprise had their Claude Code 5-hour limits doubled. The free tier was excluded. The Claude API had a separate set of changes: Tier 1 saw a 1500 percent increase in maximum input tokens per minute and a 900 percent increase in maximum output tokens per minute, per Anthropic's announcement, but that is the API rate limit (separate accounting), not the plan-quota float that Claude Code uses on Max. If you run Claude Code via your plan, only the 5-hour Claude Code limit doubled. The weekly buckets did not move.