Anthropic runs two ceilings, not one: a rolling five-hour session window and a weekly cap that sits above it. Claude Fable 5 obeys a third rule that changes with your plan, billing as usage credits on Pro and taking up to half the weekly pool on Max. Most plan advice models none of the three. Set your real usage below and this ranks Pro, Max 5x, Max 20x, stacked accounts, usage credits, hosted inference, and a raw API key by what they cost and whether they make it to Friday.
Wall-clock time with an agent working, not time at the keyboard.
Two agents running for an hour is two agent-hours. This is the single biggest driver.
Everything reduces to one unit: the Sonnet-equivalent agent-hour. Opus, Haiku and Fable convert into it at their published API price ratios, because those ratios are the closest public proxy Anthropic gives for relative allowance burn. Opus 5 is $5 in and $25 out per million against Sonnet 5's $2 and $10, so an Opus hour counts as 2.5. Haiku 4.5 is $1 and $5, so a Haiku hour counts as 0.5. Claude Fable 5 is $10 and $50, exactly five times Sonnet and twice Opus, so a Fable hour counts as 5.
Prices and limits checked August 2026, Fable 5's plan rules on the 20th. Anthropic changes both; check the numbers before you commit a year to one.
Claude Code pricing looks like three numbers. It is actually a ladder with seven rungs, and the four that are not on the pricing page are the ones that decide most real budgets.
Every Claude plan gives you the identical models, the identical Claude Code, and the identical feature set. There is no better model at $200. What you are buying is headroom, and headroom is metered by two independent ceilings that count at the same time.
The first is a rolling five-hour session window. It is not a bucket that empties on the hour: capacity returns continuously as your older usage ages past the five-hour mark. Meeting it costs you an afternoon. The second is a weekly cap that resets on a fixed cycle tied to your account, with a separate weekly cap on Opus sitting inside it. Meeting the weekly cap on a Tuesday costs you the rest of the week, and switching from Opus to Sonnet will not clear it, because the weekly and session caps are shared across every model. Only the Opus cap is model-specific, and that one clears in the four seconds it takes to run /model.
Those two ceilings do not scale together, which is the single most misread fact in Claude pricing. Anthropic describes Max 5x and Max 20x as five and twenty times Pro per session, so per session, 20x really is four times 5x. The published weekly floors tell a different story: roughly 140 hours of Sonnet on Max 5x against 240 on Max 20x. That is 1.7x, not 4x. Doubling your spend from $100 to $200 quadruples one ceiling and adds about seventy percent to the other. If the cap you keep meeting is the weekly one, the upgrade you are about to buy is mostly not aimed at your problem.
Pro's weekly floor of about 40 Sonnet-hours works out to roughly two hours of agent work a day across a five-day week, and its five-hour window binds well before that on any day you concentrate the work. Pro is the right plan for someone who uses Claude Code as a tool a few times a day, not as the way they write code.
Max 5x at $100 is the default answer for a working engineer, and the honest reason is the comparison nobody makes: Anthropic's own documentation puts metered Claude Code usage at about $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments. The average metered developer already spends more than Max 5x costs. On a subscription the marginal cost of one more turn is zero, which changes how you work more than the money does.
Fable 5 is the Mythos-class model above Opus, and it is the one model whose economics change depending on which plan you are standing on. That makes it impossible to fold into a model mix and forget, which is why it gets its own slider above.
On Pro and Team standard seats it is not part of your plan at all. It is not in the weekly limit, it does not touch the five-hour window, and it bills as usage credits at API rates from the first token. That is better than it sounds for light use and worse than it sounds for heavy use: a couple of Fable hours a week is a genuinely cheap addition to a $20 plan, and a couple of Fable hours a day is a metered bill that quietly passes what Max costs.
On Max 5x, Max 20x and Team premium seats it is included, at up to 50 percent of your weekly limits. Read that carefully, because it is a share and not a grant: Fable draws from the same weekly pool as everything else, it may take at most half of that pool, and at $10 in and $50 out per million it burns that pool about twice as fast as Opus and five times as fast as Sonnet. Half a Max 20x week spent on Fable is roughly 24 Fable-hours, not 120. The upside of the cap being Fable-specific is that meeting it never locks you out of Claude: at least half your allowance is still sitting there for Opus and Sonnet.
The crossover is sharper than most model choices. Below roughly two and a half Fable-hours a week, Pro plus credits is the cheapest way to run Fable. Above that, the $100 plan wins and keeps winning until the 50 percent cap becomes the binding constraint rather than the price.
Most comparisons treat $100 and $200 as a binary. Since usage credits shipped it is not one. On Pro or Max you can turn credits on, set a monthly spending cap, and let anything past the included allowance bill at standard API rates. That converts an occasional two-week crunch from a permanent $100 upgrade into a metered top-up you can watch. The arithmetic is simple: if your credit spend regularly clears about $100 a month, 20x is cheaper and involves no thinking; below that, especially if your heavy weeks are seasonal, 5x plus a capped credit budget wins.
Two stacked Max 5x accounts cost the same $200 as one Max 20x. One Max 20x raises one five-hour window and one weekly cap. Two accounts give you two of each, which is a genuinely different shape: when one account meets a cap, the other has an untouched week. For parallel-session work that is usually the better $200, and the reason it is rarely recommended is that nothing in the Claude CLI moves a session between accounts, so you have to notice and switch by hand.
The same argument runs one rung higher and almost nobody makes it. Two stacked Max 20x accounts at $400 is a real plan, and for the heaviest profiles it is the only Anthropic answer that survives: eight hours a day, several parallel agents, Opus or Fable as the default. It buys two independent five-hour windows, two independent weekly pools, two Opus floors, and two Fable half-shares, which is roughly 480 Sonnet-equivalent hours a week against one Max 20x's 240. Nothing Anthropic sells above it exists as a subscription, so the honest alternatives at that level are metered credits, an API key, or hosted inference. The picker will put you here when your profile earns it, and it names what the cheaper $200 options failed at so you can check the upgrade is aimed at your actual problem.
The operational cost is real and worth stating plainly: two subscriptions is two logins, two invoices, and two sets of gauges to watch, with no routing between them out of the box. Continuum runs several accounts side by side with a live gauge on each, which turns the switch into a click rather than a re-login, but the two bills are still two bills.
An API key has no session window and no weekly cap. Nothing ever stops, which is the point and also the risk. It is the correct tool for CI, scheduled jobs, and anything that runs without a person present, and it is the standard escape hatch for a week you have already burned. Set a workspace spend limit in the Claude Console before the first session rather than after the first surprise.
Hosted inference is the other shape: a weekly dollar allowance rather than an hour window. Continuum's Plus, Max 100, Max 200 and Ultra plans give you $25, $100, $200 and $1,000 of inference a week on any model, through an OpenAI- and Anthropic-compatible endpoint, with no five-hour window and no per-model cap. Against two stacked Max 20x accounts the trade is legible in both directions: one bill and one login instead of two, and a ceiling denominated in dollars you can read rather than hours you have to infer; but a subscription's marginal turn is free once bought, while an allowance is spent as it goes. Those plans are in the ranking above on the same rule as everything else, and they are never allowed to displace a cheaper Anthropic plan that also survives your week.
Prices read from Anthropic on 19 August 2026, Fable 5's plan rules on the 20th. Weekly hours are Anthropic's own published estimates, shown as the ranges they were published as.
| Plan | Price / mo | Sonnet hours / wk | Opus hours / wk | Fable 5 | 5-hour window | Best for |
|---|---|---|---|---|---|---|
| Claude Pro | $20 | 40 to 80 | not published | usage credits, outside the plan | baseline | A few agent hours a day. |
| Claude Max 5x | $100 | 140 to 280 | 15 to 35 | up to 50% of the pool | 5x Pro | One engineer, one session, all day. |
| Claude Max 20x | $200 | 240 to 480 | 24 to 40 | up to 50% of the pool | 20x Pro | Parallel sessions and Opus by default. |
| Max 5x plus credits | $100 plus overflow | 140 to 280, then metered | 15 to 35, then metered | 50% of pool, then metered | 5x Pro, then metered | Seasonal crunches, not constant load. |
| Two stacked Max 5x | $200 | 2 × 140 to 280 | 2 × 15 to 35 | 50% of each pool | two windows | Parallel work that keeps meeting weekly caps. |
| Two stacked Max 20x | $400 | 2 × 240 to 480 | 2 × 24 to 40 | 50% of each pool | two windows | The heaviest profiles: all day, parallel, Opus or Fable. |
| Anthropic API key | metered | no cap | no cap | metered | none | CI, automation, an already-burned week. |
| Continuum Plus | $25 | $25 a week of hosted inference, any model, no hour window and no per-model cap | Hosted overflow beside your own plans. | |||
| Continuum Max 100 | $100 | $100 a week, same terms | One plan instead of a plan plus credits. | |||
| Continuum Max 200 | $200 | $200 a week, same terms | Max 20x-shaped load without the caps. | |||
| Continuum Ultra | $500 | $1,000 a week, same terms | The two-stacked-Max-20x profile, on one bill. | |||
Anthropic publishes the weekly figures as estimates that vary with codebase size and session concurrency. The picker above uses the low end of each range, which is the number worth budgeting against. Fable 5's 50 percent share is a cap on how much of that same weekly pool Fable may take, not a second allowance beside it, and Fable spends the pool about twice as fast as Opus does.
Longer versions live in the Claude Code pricing guide and Max 5x vs 20x.
Max 5x at $100 is worth it the moment Claude Code is your main way of writing code, because Anthropic's own metered average is $150 to $250 per developer per month and Pro's weekly floor of roughly 40 Sonnet-hours runs out around two hours a day. Max 20x at $200 is worth it only when you run parallel sessions, use agent teams, or default to Opus, because 20x raises the same two ceilings rather than removing them.
Claude Code refuses further requests and names the reset day. The weekly cap is shared across every model, so switching from Opus to Sonnet does not restore access. Your three options are waiting for the reset, turning on usage credits so overflow bills at standard API rates against a spending cap you set, or moving that week's work onto a second account or an API key.
The five-hour window is a rolling session allowance that refills continuously as older usage ages out, so it costs you an afternoon at most. The weekly cap sits above it and resets on a fixed weekly cycle tied to your account, so meeting it on a Tuesday costs you the rest of the week. Both count at the same time, which is why one heavy fan-out can exhaust the week before the session window has even reset.
Per session, yes: the multipliers are against Pro, so 20x is four times 5x inside a rolling five-hour window. Weekly, no. Anthropic's published weekly floors are roughly 140 Sonnet-hours on Max 5x and 240 on Max 20x, which is 1.7x rather than 4x, and the Opus floors move from roughly 15 hours to 24. That compression is why a $200 account can still be told it hit a weekly limit.
For one supervised Sonnet session most of the working day, Max 5x at $100 is almost always cheapest, because the same work metered at API list rates prices at several hundred dollars a month. For heavier profiles the cheapest option stops being a single plan: two stacked $100 accounts give you two independent windows and two independent weekly caps for the same $200 as one Max 20x, which raises one of each.
Yes. Each Claude account carries its own five-hour window, its own weekly cap, and its own Opus cap, so two accounts genuinely double both ceilings rather than raising one. The catch is that nothing in the Claude CLI moves work between them, so you have to notice a cap approaching and switch by hand unless your editor does the routing for you. Continuum runs several accounts side by side with a live gauge on each.
Yes on Max 5x, Max 20x, and Team premium seats, with one condition that decides the whole answer: Fable 5 draws from the same weekly pool as every other model and may take at most 50 percent of it. It is a share of the allowance you already pay for, not extra capacity, and it is an expensive share, because Fable 5 lists at $10 per million input tokens and $50 per million output, roughly double Opus 5. On Claude Pro and Team standard seats Fable 5 is not in the weekly limit at all: it bills as usage credits at API rates from the first token. It is not available on the Free plan.
It turns on how many Fable hours you actually run. Below roughly two and a half Fable-hours a week, Claude Pro at $20 plus metered usage credits is cheapest, because you pay for the Fable tokens and nothing for a ceiling you are not using. Above that, Max 5x at $100 wins, because Fable comes out of the weekly pool at no extra charge up to half of it. Past half your weekly pool the cap binds rather than the price, and the real options are a second account, credits on top of the plan, or hosted inference with a weekly dollar allowance and no per-model cap. The picker above prices all four against your own numbers.
Every number you just typed is a guess about your own behaviour, and people guess badly about it in both directions. Continuum reads the session files already on your machine, prices every token against the current rate card, and shows what each account actually consumed by repo, by model, and by day. Next month you pick the plan from data rather than from vibes. For the other major subscription, use the Codex plan picker. For API fleets, model concurrency with the rate-limit headroom planner.
one command · reads local history · sends nothing on its own