Claude Fable 5: price and when to use it

What does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on the Claude API. Those are Anthropic's published list rates, read from the pricing page on 3 August 2026. The API model ID is claude-fable-5. It became generally available on 9 June 2026 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

[
  {
    "label": "Claude Fable 5",
    "input": "10",
    "output": "50",
    "cache_read": "1",
    "batch_input": "5",
    "batch_output": "25"
  },
  {
    "label": "Claude Opus 5",
    "input": "5",
    "output": "25",
    "cache_read": "0.50",
    "batch_input": "2.50",
    "batch_output": "12.50"
  },
  {
    "label": "Claude Sonnet 5 (intro rate)",
    "input": "2",
    "output": "10",
    "cache_read": "0.20",
    "batch_input": "1",
    "batch_output": "5"
  },
  {
    "label": "Claude Sonnet 5 (from 1 Sep)",
    "input": "3",
    "output": "15",
    "cache_read": "0.30",
    "batch_input": "1.50",
    "batch_output": "7.50"
  },
  {
    "label": "Claude Haiku 4.5",
    "input": "1",
    "output": "5",
    "cache_read": "0.10",
    "batch_input": "0.50",
    "batch_output": "2.50"
  }
]

Read that as a ladder. Fable 5 lists at twice Claude Opus 5, five times the current Claude Sonnet 5 rate, and ten times Claude Haiku 4.5. The ratio is the same on both sides of the bill, because every model in that chart prices output at exactly five times its own input rate. So if you know the input and output token split for a job, one price converts into any other by multiplication.

One date changes the arithmetic. Sonnet 5 is on introductory pricing of $2 and $10 per million tokens through 31 August 2026. From 1 September 2026 it lists at $3 and $15. Fable 5's own price does not change, so the gap between the two narrows from five times to a little over three times on that day. If you are writing a budget for the autumn, use the later Sonnet numbers.

The discounts, and the one multiplier that goes the other way

Prompt caching is the largest lever. A cache read costs one tenth of the base input price, which is $1 per million tokens on Fable 5. Writing to the cache costs more than an ordinary input token: 1.25 times base for the five minute cache, and 2 times base for the one hour cache. A five minute cache pays for itself after a single read, and a one hour cache after two reads. That arithmetic is the whole argument for working out where prompt caching breaks even before you switch it on everywhere.

The Batch API takes 50% off both sides, so batched Fable 5 work costs $5 and $25 per million tokens. Batch requests are asynchronous, so this only helps work that does not need an answer now. The two discounts stack, which makes a cached, batched job cheap on both axes.

The 1M token context window is included at standard rates on Claude 4.6 and later models. There is no long context surcharge, so a 900k token request bills at the same per token rate as a 9k one.

The multiplier that moves the bill up is data residency. Setting inference_geo: "us" to keep inference inside the United States applies a 1.1 times multiplier to every token category, including cache reads and cache writes. Leave the default global routing in place unless a contract requires otherwise.

One billing rule is specific to this model. Fable 5 ships safety classifiers that can decline a request. When that happens the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response rather than an error, and you are not billed for a request refused before any output is generated. If you retry the same prompt on another Claude model, fallback credit refunds the prompt cache cost of the switch, so you do not pay to warm the same cache twice.

Which plans include Claude Fable 5?

As of 3 August 2026, Anthropic's Claude Fable page says the model is available to Pro, Max, Team and Enterprise users. The free plan is not named. For developers it is generally available on the Claude API and through the cloud marketplaces listed above.

Available on a plan is not the same as included without limit. The plan comparison table on Anthropic's pricing page gates Fable behind a share of weekly usage limits on some seats and behind usage credits on others, and that arrangement moved more than once during July 2026. Anthropic's own statement on redeploying Fable 5 included the model for up to 50% of weekly usage limits through 7 July 2026 on Pro, Max, Team and select Enterprise plans, with usage credits after that date. Reporting through the rest of July described further extensions and a split between seat types.

So this page will not print a per plan limit. No figure published during that month held for a fortnight, and a stale number here would be worse than none. Open the row labelled Fable in the plan comparison table on the day you decide, and treat any number in a news article as out of date.

What is stable is the API rate. On every route, once an included allowance is used up, further Fable 5 usage bills at the prices in the chart above. Enterprise makes that structure explicit, because the Enterprise seat fee buys access while the tokens themselves still meter at API rates on top of it. That makes the price list the number to plan against either way, which is the same conclusion you reach when comparing API billing against a subscription for any other Claude model. If you have not picked a tier yet, start from which Claude plan matches your usage.

Why your bill runs higher than the price ratio suggests

Two documented mechanisms make Fable 5 cost more than a straight price comparison implies.

The first is the tokenizer. Fable 5 uses the tokenizer introduced with Claude Opus 4.7. Compared with models released before Opus 4.7, the same text produces roughly 30% more tokens, and the exact increase depends on the content. Haiku 4.5 predates that tokenizer, so the ten times gap in the price list is closer to thirteen times once you feed both models the same block of text. Sonnet 5 uses the newer tokenizer, so Fable against Sonnet is a fair per token comparison. Fable against Haiku is not.

The second is thinking. Adaptive thinking is always on for Fable 5, and thinking: {"type": "disabled"} is not supported. Thinking tokens are billed as output tokens, at the full output rate. The raw chain of thought is never returned on this model, and thinking.display defaults to "omitted", which means a short visible answer can carry a large billed output count. Anthropic's documentation states it plainly: the billed output token count does not match the visible token count in the response.

Read the breakdown instead of guessing. The field usage.output_tokens_details.thinking_tokens reports how many billed output tokens went to reasoning:

{
  "usage": {
    "input_tokens": 25,
    "output_tokens": 348,
    "output_tokens_details": {
      "thinking_tokens": 312
    }
  }
}

In that example, taken from Anthropic's thinking documentation, 312 of the 348 billed output tokens were reasoning nobody ever saw. When you stream, this breakdown arrives only on the final message_delta event, so a client that stops reading at the last text chunk will never record it.

The control is effort, set at output_config.effort, with levels low, medium, high (the default), xhigh and max.

{
  "model": "claude-fable-5",
  "max_tokens": 32000,
  "output_config": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "..." }]
}

Anthropic's guidance for this model is to start at high, move up to xhigh only for the most capability sensitive work, and step down to medium or low for routine work, on the grounds that lower effort on Fable 5 often beats xhigh on earlier models. Two traps come with that. Changing effort between requests invalidates your prompt cache, because the resolved effort value is rendered into the prompt, so pick a level per workload and hold it constant. And max_tokens is a hard cap on total output, thinking and response text combined, so a cap sized for an answer with no thinking will truncate. Seeing stop_reason: "max_tokens" means you either raise the cap or lower the effort.

Cost per completed task, not cost per token

[
  {
    "label": "Short answer (5k in, 1k out)",
    "fable_5": "0.10",
    "opus_5": "0.05",
    "sonnet_5": "0.02",
    "haiku_4_5": "0.01"
  },
  {
    "label": "Coding session (300k in, 40k out)",
    "fable_5": "5.00",
    "opus_5": "2.50",
    "sonnet_5": "1.00",
    "haiku_4_5": "0.50"
  },
  {
    "label": "Long agent run (2M in, 400k out)",
    "fable_5": "40.00",
    "opus_5": "20.00",
    "sonnet_5": "8.00",
    "haiku_4_5": "4.00"
  }
]

Those 3 rows are arithmetic, not a benchmark. The rates are published. The token volumes are assumptions you should replace with figures from your own usage data. A coding session shaped like the middle row costs $5.00 on Fable 5 against $1.00 on Sonnet 5. The long run costs $40.00 against $8.00. The Haiku column in that last row is arithmetic only: Haiku 4.5 has a 200k token context window, so it cannot hold that job at all.

Cost per token is the wrong unit for a decision. Cost per completed task is cost per attempt multiplied by the number of attempts. At today's rates one Fable 5 attempt costs the same as five Sonnet 5 attempts, so retry count alone almost never justifies the upgrade. Sonnet would have to fail more than four times in five before the cheap path loses on tokens.

What does justify it is everything the token bill leaves out. The gap between the two long runs in the chart is smaller than one hour of engineer time in most markets. If the expensive model saves an hour of review, or avoids one bad migration you then have to unpick, it has already paid for itself on a bill that never shows the saving. Work the comparison in money per accepted result, and include your own time in the total. Knowing what a million tokens actually buys makes that estimate far less abstract.

Three jobs where Claude Fable 5 earns its price

Long horizon agent runs where a wrong turn is expensive. Fable 5 is built for work measured in hours: a 1M token context window, up to 128k output tokens per request, and an xhigh effort level aimed at tasks running over thirty minutes with token budgets in the millions. Weigh the run against the cleanup after an agent takes a wrong branch at step 40 that nobody notices until step 300.

A migration or an audit you run once across a large codebase. One off work has no volume to amortise, so the per token premium lands on a single invoice, and the whole job may fit inside one context window. If it can run asynchronously, the Batch API halves it to $25 per million output tokens.

The task a cheaper model has already failed twice. Two failed attempts plus your debugging time cost more than one Fable attempt at the volumes in the chart above. Escalating is the cheaper branch, and it is what a model ladder is for. If you are still choosing between tiers in general, the capability comparison between the Claude model tiers answers a different question than this page does.

Three jobs where Sonnet 5 or Haiku 4.5 is the honest answer

High volume classification and extraction. These jobs are graded on throughput and unit cost, and the quality ceiling is low enough that a cheap model reaches it. At ten times the price for a task Haiku 4.5 completes correctly, Fable 5 buys nothing you can measure.

Interactive work where latency is the product. Anthropic's model table lists Fable 5's comparative latency as slower, and thinking cannot be turned off. A chat box or an editor completion feels worse on the more capable model, because users register the wait before they register the quality.

Anything that depends on events after January 2026. Fable 5's reliable knowledge cutoff is January 2026. Opus 5's is May 2026. The more expensive model is the less current one, so on recency questions you pay double for older knowledge unless you hand it search or fetch tools.

One constraint sits outside cost entirely. Fable 5 carries 30 day data retention and is not available under zero data retention, because it is designated a Covered Model. If your agreement requires zero retention, Fable 5 is not an option at any price, and no discount changes that.

What the fable-method repos show, and what they do not

Practitioners have published repositories that distil how Fable 5 approaches work into skills a cheaper model can follow. The fable-method repo describes a thinking skill, an orchestration loop, and an adversarial verifier that re-runs every claimed check rather than reading the agent's own report. Its README says so directly: "This is a community distillation, not an Anthropic artifact."

Treat it as exactly that. It is evidence of how people are driving the model, and it is not documentation of how the model works. Nothing in it is validated by Anthropic, and several near identical forks exist with different wording, which is what you would expect from folklore rather than a spec.

The signal worth taking for a cost decision is that the parts people found worth copying are procedural. Classify the task, name the check that proves it is done, gather evidence before deciding, then verify by observation instead of by reading a summary. A cheaper model handed an explicit checklist and a verification step closes part of the gap, so run that experiment on your own evaluation set before you standardise on the expensive model. The result depends on your tasks, which is why nobody else's number is worth much to you.

Check your own numbers before you commit budget

• Read usage on every response, and log output_tokens_details.thinking_tokens separately from visible output. Reasoning you never see is the line item that surprises people.
• Set effort explicitly instead of accepting the default, and hold it constant inside any conversation that relies on prompt caching.
• Put your stable prefix behind a cache breakpoint first. A cache read at one tenth of base input price saves more than most model downgrades do.
• Move anything that does not need an immediate answer to the Batch API.
• Price the alternative honestly. Comparing Claude and ChatGPT API pricing on your own workload is worth an afternoon before a long commitment.

If the workload is an agent running on your own server, the controls that matter most sit outside the model choice. Keeping agent costs under control on a VPS covers the loop limits and spend caps that stop a runaway session, and where Claude Code actually spends tokens explains the numbers you will see in your usage dashboard.

FAQ

How much does Claude Fable 5 cost per million tokens?

Claude Fable 5 lists at $10 per million input tokens and $50 per million output tokens on the Claude API, checked on Anthropic's pricing page on 3 August 2026. A cache read costs $1 per million tokens, one tenth of the base input rate. The Batch API takes 50% off both sides, giving $5 and $25 per million tokens for asynchronous work. Keeping inference inside the United States with the inference_geo parameter adds a 1.1 times multiplier to every category.

Is Claude Fable 5 included in a Claude Pro subscription?

Anthropic's Claude Fable page lists the model as available to Pro, Max, Team and Enterprise users, and does not name the free plan. Included access is not unlimited. Some seats reach Fable as a share of weekly usage limits, and others reach it through usage credits, and that arrangement changed more than once during July 2026. Read the row labelled Fable in the plan comparison table on Anthropic's pricing page on the day you decide, rather than trusting a figure from an article. Once an included allowance is spent, further usage bills at the API rate.

Why is my Claude Fable 5 bill higher than the visible output suggests?

Adaptive thinking is always on for Claude Fable 5, thinking tokens are billed as output tokens, and the raw chain of thought is never returned. With thinking.display at its default of omitted, you see an empty thinking field while paying for every reasoning token behind it. Read usage.output_tokens_details.thinking_tokens in the response to see the split, and note that when streaming it arrives only on the final message_delta event. Lower the effort level if the ratio of reasoning to answer is higher than the task needs.

Should I use Claude Fable 5 or Claude Opus 5?

Claude Opus 5 costs half as much per token and has a more recent reliable knowledge cutoff, May 2026 against January 2026 for Fable 5. Start on Opus 5 and escalate when a task fails there, or when a wrong answer costs more than the price difference. Fable 5 is the right call for long horizon agent work where one bad branch costs hours of cleanup, and for one off jobs whose premium lands on a single invoice instead of on every request forever.