Which models Claude Max gives you

Which models Claude Max gives you

Claude Max gives you the same model families that Claude Pro reaches. The higher tier buys headroom rather than a different product: more work before you run out, and a much larger share of that work on the most capable model. Model access is set by your plan, and the strongest model is metered on its own allowance, so it runs out before the rest of your plan does.

Two rules survive every release, so learn them before you learn any model name. Your plan decides which models appear in the picker and how much of each you get. The most capable model carries the tightest allowance, which is why heavy users hit that one wall first while everything else keeps working.

The model names below are dated on purpose. They change every few months. The mechanics do not.

The model list, checked on 2 September 2026

On Anthropic's plan and pricing page, checked on 2 September 2026, the Pro, Max 5x and Max 20x rows list the same four model families: Fable, Opus, Sonnet and Haiku. Max 5x is 100 USD per month and Max 20x is 200 USD per month at that date. The gap between them is allowance, not catalogue.

On Anthropic's model overview, checked the same day, the current lineup is Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. Older builds including Claude Fable 5, Claude Opus 4.8 and Claude Sonnet 4.6 are listed as legacy models that you can still select.

Read both pages yourself before you pay for anything. This guide will be old at some point, and the model at the top of the list today becomes a legacy entry later. Anthropic's own rows are the only source that is current by definition.

A couple of access details are conditions rather than names, and they catch people out:

• On Max, Opus with the 1 million token context window is included. On Pro the same window draws on usage credits.
• Fable requests can bill to usage credits instead of your plan's included limits, depending on your plan and seat tier. In an interactive Claude Code session you get a consent prompt before that happens.

If you are still choosing between the two Max tiers, the price and allowance comparison is a post of its own: the 5x and 20x tiers compared side by side. What a single subscription stretches across is covered in what one Claude Max plan actually includes.

Why the top model runs out first

Your subscription is metered in a rolling five-hour session window and a weekly window. Anthropic's usage guidance tells you to check when your plan's weekly usage limit resets "for Opus only and all other models", and that phrase is the whole mechanic: Opus has its own weekly counter, and every other model shares a different one.

The published list prices behind these models show why the split exists.

[
  {
    "label": "Haiku 4.5",
    "input_usd_per_mtok": 1,
    "output_usd_per_mtok": 5
  },
  {
    "label": "Sonnet 5",
    "input_usd_per_mtok": 2,
    "output_usd_per_mtok": 10
  },
  {
    "label": "Opus 5",
    "input_usd_per_mtok": 5,
    "output_usd_per_mtok": 25
  },
  {
    "label": "Fable 5.1",
    "input_usd_per_mtok": 10,
    "output_usd_per_mtok": 50
  }
]

Those 4 models are the current lineup at Anthropic's published API list prices, checked on 2 September 2026. Opus 5 output lists at 25 USD per million tokens, against 10 USD for Sonnet 5 and 5 USD for Haiku 4.5. Fable 5.1 is the most expensive entry at 50 USD, and what Fable is actually worth using for explains when that price earns itself. A flat allowance spent entirely on the top model would be worth several times the same allowance spent on Sonnet, so the expensive model gets counted on its own. This is also why matching the model to the task beats leaving the strongest one selected by default: which model to reach for on which job goes through it case by case.

What Claude Code prints when the allowance is gone

Claude Code has four different limit messages, and they do not mean the same thing.

• You've hit your session limit and You've hit your weekly limit are plan-wide. They count across every model, so switching model with /model does not restore access.
• You've hit your Opus limit and You've hit your Sonnet limit apply only to that model family. Switching to a model outside that family with /model keeps you working.

That distinction answers the question most people are really asking. Claude Code does not quietly move you to a lighter model when your Opus allowance is gone. It stops, prints the message, and tells you when the window resets. You then pick one of these:

• Run /model and choose a model outside the exhausted family. Sonnet or Haiku still work while the Opus counter is empty.
• Wait for the reset. On Claude Code v2.1.234 and later the session can wait and then continue the interrupted task on its own; when it does not start that wait itself, you pick it from /rate-limit-options.
• Turn on usage credits with /usage-credits and keep going past the plan allowance at a per-token price.

Run /usage before you hit a wall rather than after. It shows your plan usage bars, and you can press d or w to switch between the last 24 hours and the last 7 days. The shape of the windows and what counts against them is in how Claude usage limits are measured. When each window resets and what your options are once you have hit one are both worth reading once, so the timing stops surprising you.

How to notice you are being served a different model

Claude can switch models on its own in two situations, and neither one is about running out of allowance.

The first is a safety classifier. On Claude Opus 5, a request that the cybersecurity classifier flags is re-run on Claude Opus 4.8 instead, and Anthropic's help page says "You'll see a notice explaining that the model switched, and the response will be labeled with the model that answered." A biology flag on Opus 5 ends in a refusal instead of a fallback. On Fable 5.1 and Fable 5, a biology flag re-runs on Opus 5 and a cybersecurity flag re-runs on Opus 4.8. You can switch back from the model picker, and you can turn the automatic switch off in Settings under Capabilities, which makes the conversation pause instead.

The second is an availability fallback that you configure yourself. A fallback chain sends the request on to the next model when the first one will not take it:

claude --fallback-model sonnet,haiku

The same chain in settings.json:

{
  "fallbackModel": ["claude-sonnet-5", "claude-haiku-4-5"]
}

Claude Code tries each entry in order until one accepts the request. That is useful when your first choice is overloaded, which is a separate failure with its own error text: what the overloaded error means and how to clear it. It also means the model answering you is not always the model you selected, so check instead of assuming. Run /model to see which models your account can reach and which one is selected now. In the chat app, read the model label on the response itself.

What a model change costs you mid-task in Claude Code

Switching with /model changes which model answers your next request. The conversation history stays where it is, so you do not lose the session or the files already read. What changes is judgement, and that shows up in the middle of multi-step work.

A lighter model picks up the same context and often takes a different route through it, re-reading files the previous model had already understood, or writing a simpler fix than the plan called for. That is not a fault. It is a smaller model doing its best with a plan it did not write.

Two habits make this cheap. Switch at a task boundary rather than halfway through a multi-file edit, so the new model starts on a clean step. And use the built-in split when the work suits it: the opusplan setting uses Opus while you are in plan mode and switches to Sonnet automatically for execution, which is the same trade made on purpose instead of by accident.

There is one more reason the Opus allowance disappears faster than your typing suggests. Claude Code sends the whole conversation with every request, and every batch of tool results becomes another request carrying that history. A one-line question in a session you have had open all day still draws usage for the entire conversation. Your first message after a long break also misses the prompt cache, so the full context is reprocessed at full price. Clearing between unrelated tasks with /clear is the largest single saving available to a heavy user.

When the API beats a bigger subscription

A subscription is a fixed price with a ceiling you cannot raise inside the tier. The API is an open meter with no weekly window. The decision point is the shape of your work rather than its volume.

Stay on Max while the work is interactive and you are the bottleneck. One person typing and reading every answer rarely out-spends the monthly price. Anthropic's Claude Code cost documentation, checked on 2 September 2026, puts enterprise usage at around 13 USD per developer per active day and 150 to 250 USD per developer per month, with 90% of users below 30 USD per active day. Those are published figures for organisations paying per token, and they land close to what Max 20x costs.

Move to the API when the work runs without you: scheduled jobs, batch pipelines, many agents in parallel, or per-project cost attribution that one subscription cannot produce. A plan cannot tell you what a single client's work cost. A separate API key per project can.

Usage credits sit between the two, and they carry one property worth knowing before you lean on them. The prompt cache lifetime is one hour on a subscription and drops to five minutes once you are drawing on usage credits, so a long session costs more per request in that mode: the cache goes cold between your messages and the context is reprocessed. The full side-by-side with numbers is in the API against subscription cost comparison.

For readers in Taiwan and Hong Kong, add a note on price and a note on timing. Plan prices are listed in USD, and what lands on your card depends on local billing and tax, which is covered in what Claude Max costs in Taiwan and Hong Kong. The reset windows follow your account rather than your office hours, so a weekly reset can fall in the middle of a Tuesday afternoon in Taipei. The limit message states the exact reset time. Read it instead of guessing, and schedule the heaviest Opus work for just after it.

FAQ

Does Claude Max include models that Claude Pro does not?

On Anthropic's pricing page, checked on 2 September 2026, Pro, Max 5x and Max 20x all list the same four model families: Fable, Opus, Sonnet and Haiku. The difference between the plans is allowance rather than catalogue, with a few conditions attached. On Max, Opus with the 1 million token context window is included, while on Pro that window draws on usage credits. Model availability changes between releases, so read the plan comparison at claude.com/pricing before you buy rather than trusting a list you read months ago.

What happens in Claude Code when I run out of Opus?

Claude Code prints You've hit your Opus limit along with the time the window resets. It does not move you to another model on its own. The Opus and Sonnet limits apply only to that model family, so running /model and choosing a model outside the family lets you carry on immediately. The plan-wide messages behave differently: You've hit your session limit and You've hit your weekly limit count across all models, and switching model does not help there.

How do I know which model actually answered me?

In the chat app, a response that came from an automatic fallback is labeled with the model that answered, and a notice explains why the switch happened. In Claude Code, run /model to see which models your account can reach and which one is selected, and run /usage for your plan usage bars, pressing d or w for the last 24 hours or the last 7 days. If you configured a fallback chain with --fallback-model or the fallbackModel setting, remember that the reply can come from the second entry whenever the first is unavailable.

When is the API a better answer than upgrading to Max 20x?

Upgrade the subscription while the work is interactive, because one person reading every answer rarely out-spends the monthly price at current list rates. Move to the API when the work runs without you: scheduled jobs, batch pipelines, parallel agents, or billing that has to be split per project. Usage credits are the middle option, with one catch to plan around: the prompt cache lifetime drops from one hour to five minutes while you draw on them, so long sessions cost more per request in that mode.