Skip to content
The Frontier Brief
Release alert
Weekly guide / updated Oct 31, 2026 · written for Fable 5.1 and Opus 5.5

How to spend a week on Claude: the $20, $100 and $200 plans

The three plans do not differ in whether you get Fable. They differ in how many days Opus 5.5 at high effort lasts before you have to switch models, and in what to pair it with. Updated weekly.

This edition

Where each line on this page comes from

Updated
Oct 31, 2026
Written for
Fable 5.1 and Opus 5.5, the current Claude models on that date. When Anthropic ships a new model, the numbers are measured again.
Our experience
The lines under each plan, in the upgrade ladder and in the token tips. They come from our own paid accounts, as of the date at the top of this page. These are measured impressions, not Anthropic’s numbers.
What people say online
The section What people are saying online: 41 dated quotes collected Oct 31, 2026, each linked to where it was said. We report them; we do not check them.
Official
The lines marked Official rule come from Anthropic’s own pages, read Nov 5, 2026. Prices and plan names are Anthropic’s. Our usage-limits page has more. Claude plans (opens in a new tab)
Pro · $20 a month

Opus 5.5 on high, mixed with medium, lasts three or four days.

What a week looks like ↓Max · $100 a month

Opus 5.5 on high all week, with a little Fable 5.1, works very well.

What a week looks like ↓Ultra · $200 a month

Opus 5.5 on high from Monday to the weekly reset.

What a week looks like ↓
Pro · $20 a month

Opus 5.5 on high, mixed with medium, lasts three or four days.

  • Opus 5.5 at high effort, mixed with medium effort, lasts three or four days.
  • Add DeepSeek 4.1 Flash through OpenCode Go ($10): select DeepSeek alone, and have it carry out the instructions Opus writes until the job is finished, not before.
  • GLM is not recommended: it is weaker than DeepSeek 4.

Official rule: On Pro, Fable is not part of the plan; it runs on usage credits. The rule, in Anthropic’s words, is on our usage-limits page. Fable weekly limit and usage credits →

Max · $100 a month

Opus 5.5 on high all week, with a little Fable 5.1, works very well.

  • About 4 billion tokens a week.
  • Opus 5.5 at high effort all week, with only a little Fable 5.1, is already very good.
  • Have the main task hand work down to Opus 5.5 on medium and Sonnet 5.5 on high; Claude Code lets you mix models freely.

Official rule: On Max, Fable 5 and 5.1 may use up to 50% of your weekly limit. That is a cap inside the same limit, not an extra allowance. Anthropic says Max gives five times Pro’s usage every five hours and every week. The 50% cap, explained →

Ultra · $200 a month

Opus 5.5 on high from Monday to the weekly reset.

  • Opus 5.5 at high effort from Monday until the weekly reset.
  • It feels like 7–15 billion tokens a week.
  • Roughly 1.7–2.0× the $100 plan; estimates online range from 1.7× to 3.4×.
  • In API terms, about $12,000 a month, or $3,000 a week.
  • With every 5-hour window used to the full, you get through 5–7 of them in the week.

Official rule: Ultra is the new name for Max 20x, still $200 a month. Anthropic says it gives 10 times Pro’s usage every week, with no five-hour limit for verified accounts. Anthropic: What is the Ultra plan →

The upgrade ladder

  • Start on $20 and use it up in a day or two. Upgrade to $100 and use that up in a day and a half. Upgrade again to $200 and run the remaining days on a fresh weekly allowance.

Anthropic says that when you upgrade from Pro to Max, your usage starts fresh on the Max limits right away and the billing cycle restarts that day. For Max to Ultra it does not say whether the week’s usage so far is cleared, so that last step is our experience, not a stated rule. See the account-safety guide. Upgrading to Max (Anthropic) (opens in a new tab) The Ultra plan (Anthropic) (opens in a new tab)

How to spend fewer tokens on any plan

  • Do not keep opening new sessions, and do not keep switching to markdown.
  • Turn on auto compact and bring the threshold down to 30–40%. Baoyu says 30%; our own setting still compacts only past 50%.
  • Caching: on a subscription, Claude Code’s main conversation now keeps its prompt cache for 1 hour instead of 5 minutes; subagents still get 5 minutes.
  • Structure: one main task, subtasks spun off from it, and one-off subagents, so the main context stays small. Two weeks ago, opening a conversation ate 20% of a 5-hour window; that has since been optimized away.
  • Use cloud sessions and let them work with agents on their own; we show examples.

Anthropic’s own list of ways to use limits well is short and official. Usage limit best practices (Anthropic) (opens in a new tab)

Online · reported, not checked · collected Oct 31, 2026

What people are saying online

Other people’s accounts of the same three plans, in their own words, each with its date and a link to where it was said. We record what was said; we do not check it, average it or turn it into a number of our own. Where it disagrees with our lines above, both stay.

Pro, $20

  • GitHub, Fei2411/real-api-pricing issue #124 (ganpei777, Claude Pro, Opus 5.5) · Oct 18, 2026 “Dividing the 0%→100% total by 100 gives $12.66 per weekly percentage point, and $1,266.15 for this full week.” API-list-price equivalent of one Pro week on Opus 5.5; the repo maintainer notes that cache reads were 98.96% of it. Source ↗
  • GitHub issue #104967 (Fast-and-Pythonic, Pro, Opus 5.5 and Sonnet 5.5) · Oct 25, 2026 “So in the plan’s limit, a Sonnet token seems to cost about 0.9 of an Opus token, while at API prices it costs half.” Same post: “one point of the weekly window cost $4.45 at API prices in mid-September and $2.31 in mid-October.” Source ↗
  • GitHub issue #101937, comment (sergiokapone, Pro, Opus 5.5) · Oct 27, 2026 “Windows then lasted ~3.5 h before the limit; now ~1.5 h.” Same comment: “API-equivalent cost of the whole window: roughly $13-14.” The window is the 5-hour one. Source ↗
  • GitHub issue #105352 (lebiraja, Pro, Sonnet 5.5 and Opus 5.5) · Oct 27, 2026 “About 50% of my weekly limit was used in two sessions over about two days. A month ago the same style of work lasted me the whole week.” The author later adds: “I did not realise some of my custom agents were pinned to Opus.” Source ↗
  • GitHub issue #69904, comment (TheDeveloper404, Pro at €20) · Jul 1, 2026 “Weekly limit was at 0% Saturday 4PM. By Wednesday — before using Claude Code at all that day — weekly was already at 100%.” Model generation not stated; mid-2026. Source ↗
  • GitHub issue #46829, comment (rinchen, Pro) · Apr 12, 2026 “This perfectly explains why it would take hours before I ran out of Claude on Pro to now running out in 20 minutes.” Said in the thread about the prompt-cache TTL dropping to 5 minutes (Sonnet 4.6 / Opus 4.6 era). Source ↗
Seen in search results only: 2 for this plan

These were visible as search snippets but could not be opened from the session that collected them (Hacker News, Reddit, Medium and X are not reachable there). Wording is approximate, the post date is inferred, and the date shown is when we saw the snippet. We keep them because they are the most-quoted claims; they go into the list above once someone opens the original.

  • Hacker News comment in the Kimi K3.1 Pro thread (username not shown) · seen Oct 31, 2026 · posted around October 2026; thread date inferred, not confirmed I tried the $20 Claude Plan and was able to reach the 5 hour limit in maybe one prompt. Another commenter: a single 5-hour allowance on the €20 plan equals about $5–10 of DeepSeek 4.1 Flash API usage. (paraphrase) Link ↗
  • Medium (Kunal Bhardwaj), “I was burning through Claude Code’s weekly limit in 3 days” · seen Oct 31, 2026 · posted around May 2026; inferred from the search result, not confirmed Every single week, without fail, I’d hit my token limit by Wednesday. The fix described: rules in CLAUDE.md that hand tests, config and docs to a cheaper outside model. (paraphrase) Link ↗

Max, $100

  • GitHub issue #99204 (nooneee-droid, Max 5x, Opus 5.5 on medium) · Oct 3, 2026 “Max 5x: weekly limit drains ~15%/day on Opus 5.5 (medium) with light use; Fable bar rises without Fable use.” The issue title. The body adds that the previous week, on Fable at max thinking with many subagents, “lasted the entire week”, and this one was “roughly 650M total tokens”. Source ↗
  • GitHub issue #97997, comment (VT1962, Max 5x, Opus 5.5 only) · Oct 1, 2026 “Last week, the Fable meter reached 100% despite my not selecting Fable.” Readings given: Sept 29, Fable 4% / weekly 8%; Oct 1, Fable 60% / weekly 62%. Source ↗
  • GitHub issue #79773, comment (erpland, Max 5x then 20x) · Sep 15, 2026 “About 10 % in the first ~10 hours on 20x, versus about 60 % in ~41 hours on 5x the weekend before.” The commenter notes the +50% → +25% promotional adjustment at the time muddies the comparison. Source ↗
  • GitHub issue #45756 (molu0219, Max 5x, Opus 4.6 with 1M context) · Apr 9, 2026 “Pro Max 5x Quota Exhausted in 1.5 Hours Despite Moderate Usage. Background sessions consumed 78% of post-reset quota.” Title plus one line from the body. Spring 2026, Opus 4.6 era. Source ↗
Seen in search results only: 2 for this plan

These were visible as search snippets but could not be opened from the session that collected them (Hacker News, Reddit, Medium and X are not reachable there). Wording is approximate, the post date is inferred, and the date shown is when we saw the snippet. We keep them because they are the most-quoted claims; they go into the list above once someone opens the original.

  • Hacker News comment in the Sonnet 5.5 release thread (username not shown) · seen Oct 31, 2026 · posted around September 2026; thread date inferred, not confirmed “Probably a first world problem, but with Opus 5.5’s efficiency, the limits on the 5x plan are simply sufficient for my everyday work, even when running 2-3 sessions at a time.” Link ↗
  • Botfarm blog (Alex Dobrushskiy), “Claude Max $100 vs $200: measured, not guessed” · seen Oct 31, 2026 · measured late February to early March 2026; publication date not confirmed One full 5-hour window on the $200 plan uses about 15% of the weekly budget, versus about 8% on the $100 plan: 4x the burst capacity but only ~2x the sustained weekly budget. A companion post puts a week at about $523 of API usage on Max 5x and about $1,100 on Max 20x. (paraphrase) Link ↗

Ultra (formerly Max 20x), $200

  • GitHub issue #99330 (Fuduntu, Max 20x) · Oct 3, 2026 “Max 20x, weekly limit hit on 2026-09-30 after ≈ $1,827 since the weekly reset. That is about 3.6 full sessions per week.” The same post’s table: “Week ≈ 1.6× (2× at most)” the 5x week; “5-hour session ≈ 3.2× (advertised 4×)”. Dollar figures are the author’s API-list-price equivalents. Source ↗
  • GitHub issue #99330, comment (Clarelive, mostly Opus) · Oct 30, 2026 “Using 21% of one 5-hour window cost 6% of the weekly limit. So one full 5-hour window costs about 29% of the week, and the whole weekly allowance is only about 3.5 five-hour windows (2.9x-4.3x allowing for rounding).” Source ↗
  • GitHub issue #79773, comments (Yoshinobu-Suzuki, 5x and 20x, Opus 5 / Sonnet 5 / Fable 5.1) · Sep 6, 2026 “My two full 5x weeks were 1.89B / 1.59B (avg 1.74B), so 20x ≈ 2.0× of 5x, not 4×.” Earlier, on July 28: “What I measure is 2x, which is exactly the ratio of the two prices ($100 to $200), not the ratio of the two tiers.” Source ↗
  • GitHub issue #79773, comment (cheonnamsu, 5x then 20x, includes Fable) · Aug 19, 2026 “Against the Max 5x maximum of $3,222 (never hit cap), the ratio is 1.58×, versus the expected 4× tier ratio.” Measured during the temporary +50% weekly-limit promotion in August 2026. Source ↗
  • GitHub issue #79773, comment (jc-myths, 5x then 20x, Opus 5.5) · Oct 22, 2026 “So with the same upgrade path and the same kind of work, one full 5-hour window now costs about 40% of the weekly limit instead of about 27% two days earlier.” Source ↗
  • GitHub issue #105437 (Allenfang2026, Max 20x) · Oct 27, 2026 “This week I had already used 76% of my weekly limit by Tuesday, only four days after the Friday reset. Meanwhile my current 5-hour session bar is at just 4%, so the weekly cap is what actually limits me, not the session cap.” Source ↗
  • GitHub issue #103376 (EthanRCT, Max 20x, Opus 5.5 main with Fable 5.1 and Sonnet 5.5 subagents) · Oct 19, 2026 “On Claude Max (20x), the weekly “all models” meter went from 0% to 54% in the first 22 hours after the weekly reset.” A commenter (justanotheruseer, Oct 26) in the same thread: “The allowance went from ~5.2M per 1% (Oct 5–9) to ~3.85M (Oct 19 week).” Source ↗
  • GitHub issue #105122, comment (AutoHenry, Max 20x, mostly Opus 5.5 and Sonnet 5.5) · Oct 27, 2026 “Oct 6 to 13 window: ~$2,605 / 11B tokens. Current window (Oct 20 to 27): ~$1,776 / 6B tokens, and /usage already shows 93% of the weekly limit used. So roughly a third less work is getting me to the cap compared with two weeks ago.” Dollar figures are the commenter’s own local logs priced at API rates. Source ↗
  • GitHub issue #105162 (ggotye, upgraded Max 5x → 20x on Oct 15) · Oct 26, 2026 “So the measured weekly capacity on Max 20x (~10–13M output tokens) is lower than what was consumed in a full week on Max 5x without ever being blocked (~14.7M).” The 20x week: “Weekly meter at 88% on Tuesday Oct 20, fully locked out Oct 21–24 (4 days).” The author believes the allowance was not recalculated after a mid-week upgrade. Source ↗
  • GitHub issue #99204, comment (timy2517, Max 20x, Opus 5.5) · Oct 18, 2026 “My weekly limit reset on Oct 17. By midday the weekly bar was already at 33%, and about 24 hours after the reset it reached 54%.” Source ↗
  • GitHub issue #100130 (youngsoo-spec, Max 20x) · Oct 7, 2026 “The weekly limit runs out 2–3 days before reset.” Source ↗
  • GitHub issue #97997 (bujawang, Max 20x, Opus 5.5 and Sonnet 5) · Sep 28, 2026 “The Fable window is 50% of the weekly limit, so 20% of it is about 10% of the whole weekly limit. So Opus 5.5 usage may be counted as Fable usage.” The author’s own workaround: “use Claude Code only with Opus 5.5 as the main model … Never select Fable, and don’t use the advisor.” Source ↗
  • GitHub issue #104750, the author’s own correction (captinjack99, Max 20x) · Oct 24, 2026 “Correction: the week of Oct 8 included a free weekly limit reset I had forgotten about, so it held about two weeks’ worth of usage. … I see no evidence of a reduced limit.” Source ↗
  • GitHub issue #69904 (danahaero, Max 20x, Opus 4.8 and 4.6) · Jun 21, 2026 “After ~3 hours of a workflow I have run for weeks, /usage showed the session at 47% but the weekly (all models) at 85%.” The /usage log in the post: “Session totals: $345.89; predominantly Opus”. Closed by bcherny on Aug 17: the allowances “come from your plan on the server side; Claude Code only displays them.” Source ↗
  • GitHub issue #50325 (Jilvert22, Max 20x, Opus 4.7 with 1M context, six workspaces at once) · Apr 18, 2026 “On the Max 20x plan, the weekly usage indicator reached ~50% after only 1.5 days of use.” Spring 2026, Opus 4.7 era. Source ↗
Seen in search results only: 8 for this plan

These were visible as search snippets but could not be opened from the session that collected them (Hacker News, Reddit, Medium and X are not reachable there). Wording is approximate, the post date is inferred, and the date shown is when we saw the snippet. We keep them because they are the most-quoted claims; they go into the list above once someone opens the original.

  • Hacker News post “Claude 20x usage is only for the 5 hour window, not for the weekly limit” (username not shown) · seen Oct 31, 2026 · posted around August 2026; inferred from the item number, not confirmed The poster estimates that Max 20x users get “roughly something like 1.7x” of Max 5x on the weekly limit. (paraphrase) Link ↗
  • Hacker News comments in the Sonnet 5.5 release thread (usernames not shown) · seen Oct 31, 2026 · posted around September 2026; thread date inferred, not confirmed One commenter burned through the 20x weekly limit in 2.5 days running a team of agents on medium reasoning; another ran out of the Opus 5.5 20x plan in 4 days out of 7. (paraphrase) Link ↗
  • Hacker News thread “Fable is now included on Max plans (up to 50% of weekly limit)” (usernames not shown) · seen Oct 31, 2026 · posted around July 2026; inferred, not confirmed One commenter says Fable “burned through all MAX 20x tokens in two days and accomplished nothing”; another says “Fable 5 ate up 10% of my weekly allowance in 30 minutes”. (paraphrase) Link ↗
  • Hacker News comment in the Opus 4.7 model-card thread (username not shown) · seen Oct 31, 2026 · posted around April 2026; inferred, not confirmed “A few months ago it was for weekly: pro = 5m tokens, 5x = 41m tokens, 20x = 83m … (the commenter adds that this may be outdated).” Link ↗
  • X post by @kimmonismus (text seen in a search result only; X not opened) · seen Oct 31, 2026 · date decoded from the post ID as 2026-08-31; not confirmed on X “Claude’s $200 Max plan offers 20x Pro usage within the five-hour window, but its weekly limit is only about twice that of the $100 plan.” Link ↗
  • ITNEXT (Jacob Ferus), “If you thought Claude Max 20x meant 20x weekly usage, you were wrong” (member-only article) · seen Oct 31, 2026 · dated 2026-09-03 in the search result; the article was not opened, so not confirmed User reports suggest the 20x weekly allowance is only around twice the total weekly allowance of Max 5x. (paraphrase) Link ↗
  • Classmethod DevelopersIO (Japanese), three-account measurement · seen Oct 31, 2026 · measured August 2026 during the +50% promotion; publication date not confirmed With one 20x and two 5x accounts, the 20x weekly limit came out at about 3.5× the 5x one (3.15–3.88×). (paraphrase) Link ↗
  • wokey.ai (a vendor of pay-as-you-go access), “Claude usage limits” · seen Oct 31, 2026 · the page says it measured the week of Oct 19–26, 2026; the page was not opened, so not confirmed A Max 20x week holds only ≈ 6.1 full 5-hour windows; about $1,602 of API usage, versus about $157 for Pro. (paraphrase) Link ↗

Any plan: compact, caching, session structure

  • GitHub issue #45756, comment (bcherny, Claude Code team) · Apr 12, 2026 “Since Claude Code uses a 1 hour prompt cache window for the main agent, if you leave your computer for over an hour then continue a stale session, it’s often a full cache miss.” The same comment recommends starting with CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 on the 1M context window. Source ↗
  • GitHub issue #102338 (leodumont444, Claude Code 2.1.300, Opus 5.5 / Fable 5.1) · Oct 15, 2026 “The new keepalive (POST /v1/messages/cache_touch every 30 s for up to 60 min of inactivity) … So the cache is kept warm for ~1h, but the status line field can no longer say when it will actually expire.” Observed by the reporter from the client’s traffic; not documented by Anthropic. A commenter (martinkozle, same day): “after 8m06s and 11m27s idle gaps, the next request read 80k / 87k tokens from cache and wrote only ~1.5k.” Source ↗
  • GitHub issue #46829 (seanGSISG, Sonnet 4.6 / Opus 4.6) · Apr 12, 2026 “Anthropic appears to have silently changed the prompt cache TTL default from 1 hour to 5 minutes sometime in early March 2026.” Replies in the thread: Jarred-Sumner (Apr 12): “The March 6 change makes Claude Code cheaper, not more expensive.” lizthegrey (May 8), quoting Anthropic: “two separate Claude Code bugs could cause eligible subscription seats to fall back to the 5-minute prompt-cache TTL”. Source ↗
  • GitHub issue #99360 (BenHamm, Opus 5.5, 16 background subagents) · Oct 4, 2026 “Every subagent cache write was ephemeral_5m_input_tokens (19.65M, 0 in the 1-hour class). Every main-session cache write was ephemeral_1h_input_tokens. … Within ~65 minutes the account hit its session limit.” “By my estimate the misses were roughly half the price-weighted cost of the run.” Source ↗
  • GitHub issue #99808 (sgaofen, mostly Opus 5.5, Claude Code 2.1.289) · Oct 5, 2026 “The Bash tool cuts the call off at about ten minutes even when a longer timeout is requested, so the subagent wakes up, finds the cache expired, and pays for the full context again.” “So a global 1h setting only saves about 5% overall, while per-type settings would save considerably more.” Source ↗
  • GitHub issue #43599 (VVisher, Opus 4.6 with 1M context, Claude Code 2.1.91) · Apr 4, 2026 “On a 500K-token session idle for 7 hours, a single message consumed 9% of the 5-hour rate limit — from 0%.” The 20–40% figure in the same post is the author’s projection for 1M+ sessions, not a measurement. Source ↗
  • GitHub issue #45756, comment (EternalRights) · Jun 18, 2026 “I left Claude Code open overnight, came back to half my quota gone with zero actual work done.” Source ↗
  • GitHub issue #45756, comment (wutlu) · Apr 12, 2026 “I solved it by resetting the session for each task and continuing with a new session.” Source ↗
  • GitHub issue #45756, comment (cnighswonger, Opus 4.6 era) · Apr 10, 2026 “Within the resolution of our data, cache_read does not meaningfully count toward the 5-hour quota.” The same commenter reported on Apr 23 that cache reads appear weighted at about 0.1x; the two readings differ. Source ↗
  • GitHub issue #82761 (Jordanm37, Opus 5 with 1M context, Claude Code 2.1.220) · Jul 31, 2026 “While the override worked — auto-compactions fire at ~32.5% of the 1M window, exactly as configured. After 2026-07-14 — zero auto events in 17 days, across all projects, despite sessions running well past the 33% mark.” About the CLAUDE_AUTOCOMPACT_PCT_OVERRIDE setting; a manual /compact is logged at 511,877 tokens (51% of the window). Source ↗
  • GitHub issue #38335, comment (tonydzi) · Oct 7, 2026 “The full listing injected into context at start was ~32,219 tokens = 31.5% of a 102,259-token session start. … Our background routines spent 71% of a weekly budget re-reading context.” The listing was 318 installed skills. Source ↗
  • GitHub issue #38335, comment (cnighswonger) · Sep 20, 2026 “The 65k cold-start floor is essentially 100% CC prompt-cache-creation.” Measured first request of a session: 64,954 tokens. Source ↗
  • GitHub issue #102248 (atashjamal5666-blip, Max, Fable 5.1 at effort xhigh) · Oct 15, 2026 “Claude launched a 137-agent Workflow for a single-file audit, against a standing “no multi-agent” instruction (~53% of weekly usage in 1 hour). ~16M tokens and ~1,475 tool calls in ~62 minutes.” Title plus one line from the body: “5-hour window went to 100%, Fable weekly allowance to 53%”. Source ↗
  • GitHub issue #101937 (Dougmayfair67, Opus 5.5, plan not stated) · Oct 14, 2026 “Despite Opus 5.5 being “cheaper and faster” since it’s been released, it burns through my entire week in the span of one to two days. Even offloaded lots of the actual work to codex to implement and it still burns through everything.” Source ↗
  • GitHub issue #105122 (casimmvancouver, Max, tier not stated) · Oct 26, 2026 “Today (Oct 25, 2026) I hit 40% of my WEEKLY limit within a few hours of normal (but sustained) work. A full week’s limit now lasts about 2-3 days on the same workflow that used to cover the week.” Source ↗
  • GitHub gist by monperrus, comment (zvorinji, Max 20x) · Jul 18, 2026 “Mine resets exactly once a week at the exact same time.” Replying to the gist’s claim that the “weekly” limit resets every 72 hours (measured June 9–20, 2026); another commenter (Headpunch, Aug 7) also says “now it resets once per week.” Source ↗
Seen in search results only: 2 for this plan

These were visible as search snippets but could not be opened from the session that collected them (Hacker News, Reddit, Medium and X are not reachable there). Wording is approximate, the post date is inferred, and the date shown is when we saw the snippet. We keep them because they are the most-quoted claims; they go into the list above once someone opens the original.

  • Hacker News comment on the May 2026 doubling of 5-hour limits (username not shown) · seen Oct 31, 2026 · posted around May 2026; inferred, not confirmed Doubling the five-hour rate limits is merely a marketing stunt if the weekly rate limit stays: it simply means you can reach the weekly limits in three days instead of five. (paraphrase) Link ↗
  • Hacker News comment (sidrag22, Opus 4.6) · seen Oct 31, 2026 · posted around May 2026; inferred, not confirmed Burns about 10% of the weekly limit in one 5-hour window; uses Opus only for planning and hands the coding to cheaper models; budgets about 14% a day, saving early in the week. (paraphrase) Link ↗

Where the accounts disagree

  • How much more a $200 week holds than a $100 week: 1.58× (cheonnamsu), about 1.6× with 2× at most (Fuduntu), 2.0× (Yoshinobu-Suzuki), and in one upgraded account less than the 5x week (ggotye). Snippets add 1.7× (Hacker News), about 2× (Botfarm, @kimmonismus, ITNEXT) and 3.5× (Classmethod). Our line above says 1.7–2.0×. Quotes: max20-cheonnamsu · max20-fuduntu · max20-yoshinobu · max20-ggotye · seen-max20-hn-1-7x · seen-max5-botfarm · seen-max20-classmethod
  • How many full 5-hour windows a Max 20x week holds: about 2.5 (jc-myths, one window ≈ 40%), about 3.5 (Clarelive), 3.6 (Fuduntu), and about 6.1 in a vendor’s snippet (wokey). Our line above says 5–7. Quotes: max20-jc-myths · max20-clarelive · max20-fuduntu · seen-max20-wokey
  • Whether the main conversation’s prompt cache lasts 5 minutes or 1 hour: “silently changed to 5 minutes” (seanGSISG), “1 hour for the main agent” (bcherny, Jarred-Sumner), “two bugs dropped eligible seats to 5 minutes” (Anthropic, as quoted by lizthegrey), and since October a keepalive that keeps it warm for about an hour while writes are logged as 5-minute (leodumont444). Quotes: any-seangsisg · any-bcherny · any-leodumont
  • Whether weekly limits shrank in mid-October 2026: yes, by about a third (AutoHenry, justanotheruseer in EthanRCT’s thread, casimmvancouver, jc-myths); no, a forgotten free reset explained it (captinjack99’s own correction). Quotes: max20-autohenry · max20-ethanrct · any-casimmvancouver · max20-jc-myths · max20-captinjack
  • Whether cache reads count against the limit: “does not meaningfully count” then “about 0.1x” (cnighswonger, April), while a Pro week in October priced at API rates was 98.96% cache reads (ganpei777) and a Pro user found a Sonnet token costing about 0.9 of an Opus token inside the limit (Fast-and-Pythonic). Quotes: any-cnighswonger-cache · pro-ganpei777 · pro-fast-and-pythonic
  • What to do after a pause: start a new session per task (wutlu); resuming an idle session replays the whole transcript uncached (VVisher); leaving it open overnight cost half a quota (EternalRights); since Claude Code 2.1.300 the client itself keeps the cache warm for up to an hour (leodumont444). Quotes: any-wutlu · any-vvisher · any-eternalrights · any-leodumont
  • Whether the Fable meter moves without Fable: three accounts say it rose to 20%, 60% and 100% with no Fable request selected (bujawang, VT1962, nooneee-droid); Anthropic’s rule is a 50% cap for Fable inside the weekly limit. Quotes: max20-bujawang · max5-vt1962 · max5-nooneee
  • Whether Opus 5.5 made the limits comfortable: “the limits on the 5x plan are simply sufficient” (Hacker News snippet) versus “it burns through my entire week in the span of one to two days” (Dougmayfair67) and 15% a day on light medium-effort use (nooneee-droid). Quotes: seen-max5-hn-sufficient · any-dougmayfair · max5-nooneee

Every line in this section has a link and a date, and the build validates both. Opened sources show the post date. Items the collecting session could not open sit in the collapsed “Seen in search results only” blocks, dated by the day we saw the snippet, with the post date marked as inferred. Posts on X appear only when a search result showed the text; we did not read X directly.

This week

First edition. From the next update on, what changed is listed here with its date.

Read next

Tell me when it ships / free

One email when it changes.

One email when Fable 5.5 is officially released, and, if you tick it, one when Skill checkout opens. The weekly guide itself stays on this page.

What to hear about

No account, no payment. How we handle your data.

From the Skill store

Video Skills people start with

All video Skills →

These four are free. Checkout is not open yet: it opens once our payment provider finishes its review, and downloads open with it. Nothing is charged until then.

  • FreeVideo

    Every Screen Agent

    An agent finishes a job on a laptop; the result shows up on a phone, then in a terminal, a chat and a code review. A 6.8-second light product shot.

    See it in the catalog →
  • FreeVideo

    Warp Liftoff

    A blue spiral galaxy, a flash and a rocket climbing off a planet, then your headline and button. An 11-second dark launch teaser.

    See it in the catalog →
  • FreeVideo

    Paint Sweep Finale

    A paint stroke wipes the frame, your wordmark is painted on letter by letter, and torn paper brings in the tagline. 8.6 seconds.

    See it in the catalog →
  • FreeVideo

    Dust Erase

    A data-broker listing is scanned, crumbles into dust, and your payoff types on in blue. A 7-second privacy-product sting.

    See it in the catalog →

Free Skills to start. Premium when you need more. Monthly $9.99 · Yearly $79 · Lifetime $169

See pricing
FAQ

Questions about this guide

Are these Anthropic’s numbers?

No. The lines under each plan are our own measured experience, as of the date at the top of this page. Anthropic’s published rules (the 50% Fable cap, usage-credit prices, the 5-hour and weekly limits) are on our usage-limits page, with links to the official articles.

What is the “What people are saying online” section?

Dated quotes from Reddit, Hacker News, blogs, GitHub and YouTube about the same three plans, each linked to its source. They are other people’s experience, reported as said; we do not check them, and they do not change our lines above.

How often does this change?

Weekly. The date at the top is the last update, and the “This week” section lists what changed.

Today: Fable 5.5 rumors