Claude Fable is Anthropic's model line for the hardest, longest-running work. The current version, Claude Fable 5.2, launched on November 9, 2026 and succeeds Fable 5.1. According to Anthropic, it is a step above Opus 5.5 in almost all the capabilities it evaluated. On the API it costs 8 dollars per million input tokens and 40 dollars per million output tokens: 20 percent less than Fable 5.1, but twice the price of Opus 5.5. On Max, Ultra and Team premium seats, Fable is included for up to half of your weekly limit; on Pro and Team standard seats it runs on usage credits. For most tasks, Anthropic still recommends Opus 5.5.
This article walks through Fable 5.2 based on Anthropic's announcement, docs and help center, as of November 13, 2026. My AI employees are defined with model: opus, so my own team runs on Opus. I have not run my own benchmarks of Fable 5.2 against Opus 5.5; every performance figure here comes from Anthropic. How Fable fits next to Opus, Sonnet and Haiku is covered in Claude Models: Opus, Sonnet, Haiku, Fable.
What is Claude Fable 5.2?
Anthropic's docs describe Fable 5.2 as built for demanding reasoning and long-horizon agentic work. The typical uses they list: agent sessions that run for hours, multistep deep research, and analysis carried all the way to a finished document, spreadsheet or deck.
The key specs according to Anthropic:
The API model ID is claude-fable-5-2. Besides its own API, Anthropic offers the model on Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and it will not retire it before November 9, 2027.
On performance, the announcement says Fable 5.2 and Mythos 5.2 are a step above Opus 5.5 in almost all the capabilities Anthropic evaluated, and that they generate output up to 30 percent faster than Fable 5.1. One example from the same page: on the FrontierCode 2.0 coding benchmark, Fable 5.2 at high effort scored about level with Opus 5.5 at max effort, for roughly half the cost per task.
What changes from Fable 5.1?
On a subscription, the switch is small. If you had claude-fable-5-1 saved in your Claude Code user settings, Claude Code moves it to the fable alias on Pro, Max and Ultra the first time you run v2.1.321 or later; on Team and Enterprise it depends on your organization's settings. After that, Fable 5.2 starts at its own default effort, medium. An old top-level level in your user settings no longer applies to Fable 5.2, so pick one again with /effort. At high and above Fable 5.2 thinks before almost every reply, and according to the docs a line like “answer directly without deliberating” doesn't stop it from thinking even at medium. If a simple answer is too slow or too expensive, lower the effort for that task. Fable 5.1 remains available on the API.
For your own applications on the API
Swapping the model ID is usually enough. Anthropic lists four changes that can break existing code:
- Only append to the conversation. If your app edits earlier messages, the system prompt or the tool list mid-conversation, the API rejects the request. If Claude Code, claude.ai, Managed Agents or the Agent SDK builds your requests, nothing changes for you.
- Reasoning stays with the model. Only Fable 5.2 and Mythos 5.2 can reuse what Fable 5.2 worked out internally. Move a conversation to another model and that reasoning is lost.
- Computer use needs the new toolset. The older computer use tool version returns an error everywhere except Amazon Bedrock.
- Advisor tool. If Fable 5.2 runs with the advisor tool, the advisor must also be Fable 5.2 or Mythos 5.2.
Anthropic also recommends re-running your evals, starting at the default effort, high.
How much does Claude Fable 5.2 cost?
On the API, per million tokens:
That puts cache reads at 2.5 percent of the input price. How API billing works and how to create a key is covered in Claude API Key: Setup, Pricing, First Call.
Fast mode on Fable 5.2
Claude Code also runs Fable 5.2 in fast mode, which Anthropic says is up to 2.5 times faster at the same quality, for double the price: $16 input and $80 output per million tokens. Three points from the docs. You need Claude Code v2.1.321 or later. On a subscription, fast mode always bills to usage credits, even if you have limits left. And it only applies if you turned it on with /fast while on Fable 5.2; switch from Opus to Fable 5.2 with fast mode on and it turns off. The rest, including the context trap when you turn it on late, is in Claude Code Fast Mode: Cost and Switching.
Fable on a subscription: limit, launch offer and Fable reset
On a subscription you don't pay per token; your usage limits are what count. Anthropic's help center says all Fable models (5, 5.1 and 5.2) work the same way:
On Enterprise, Fable 5.2 stays off until an admin enables it. Seven days after launch it is added automatically to every role that already has a Fable model, unless an admin opts out.
One thing to know on Max and Ultra: the 50 percent is a share, not an extra. You can spend up to half of your weekly limit on Fable, every other model draws from the same limit, and you never get more than your weekly limit. According to the help center, Fable also uses up your limit faster than other models. Once the Fable share is used up, you continue with usage credits or switch to another model. How Pro, Max and Ultra limits work in general is explained in Claude Pro or Max: Limits Explained.
Launch offer: a week included and Fable resets
For the launch of Fable 5.2, Anthropic set up two limited-time offers, both covering every Fable model:
Pro and Team standard seats, launch week from November 9 to 16, 2026: Fable included in your weekly limit according to Anthropic, up to half of that limit, plus one Fable reset. Both end on November 16, 2026 at 11:59 pm Pacific time. After that, Fable runs on usage credits again there, and any unused part is gone.
Max, Ultra and Team premium seats, through March 8, 2027 according to Anthropic: one Fable reset if you've subscribed for less than six months, two for six to twelve months, three for more than twelve months. The resets expire on March 8, 2027 at 11:59 pm Pacific time.
Resets only go to accounts with paid billing in place before November 9, 2026 at 10:00 am Pacific time. Subscription length counts from your first paid bill, even if you canceled in between. According to the help center, the offers don't cover the Free plan, standard seats on legacy seat-based Enterprise plans, or self-serve and usage-based Enterprise plans.
What a Fable reset does
A Fable reset does two things: it gives back the part of your weekly limit you spent on Fable, and it refills your five-hour window completely if your account has one. Usage from other models stays where it is. The help center's example: if you've used 80 percent of your weekly limit, 50 of it on Fable, you're back at 30 percent after the reset. Your limit doesn't grow, the 50 percent cap for Fable stays, and your weekly reset time doesn't move. Verified Ultra accounts have no five-hour limit, so there the reset only affects the weekly limit.
You redeem it under Settings > Usage, in the “Limit resets” section, with “Reset for free”, or straight from the message you see when you hit a limit. That works in the browser, the desktop app and the mobile apps; according to the help center it isn't currently available in Claude Code in your terminal or IDE, but the restored usage applies there too, because limits are shared across your account. You don't have to wait until you hit the limit. A used reset can't be undone, it doesn't refund usage credits you've already been billed for, and if you cancel or downgrade before using it, it's gone.
When is Fable 5.2 worth it, and when is Opus 5.5 enough?
Anthropic is clear about this. For most coding and knowledge tasks, start with Opus 5.5; Fable is meant for the most ambitious, open-ended technical work. The docs put the threshold precisely: use Fable 5.2 when Opus 5.5 at higher effort still falls short.
A simple order follows:
- Start with Opus 5.5. It is the Claude Code default on every plan and on the API. No plan starts on Fable.
- Raise the effort first. A higher effort level is a smaller step than switching models.
- Then try Fable 5.2. Especially for ambiguous problems such as root-cause investigations, outage debugging or architecture decisions, which are the examples the Claude Code docs give.
The main argument against Fable is cost: twice the per-token price of Opus 5.5, and on a subscription it uses up your limits faster than other models, according to the help center. Routine work rarely justifies it.
Tips for Fable 5.2: effort, subagents, old rules
Anthropic has published a dedicated prompting guide, “Prompting Claude Fable 5.2”. Its core message: Fable 5.2 behaves differently from earlier models in several ways, and some of what was written for Fable 5.1 or Opus 5.5 now costs more or gets in the way. These points matter most day to day:
- Choose the effort level on purpose. According to the guide, you get most of the gain at high; above that, gains are uneven and cost many more tokens. Low suits quick, well-specified tasks: on Anthropic's coding evaluations, Fable 5.2 at low came close to what Fable 5.1 scored at higher effort. You can also mix levels, for example plan at medium and switch to high for the long unattended run.
- Put subagents on another model. If neither the subagent's definition nor Claude's call names a model, a subagent in Claude Code runs on the main session's model, and its requests count toward the same usage limits. If you work with Fable, your own subagents without a model setting run on Fable too. With
model: opus or model: sonnet in the definition they run on that model instead; for all your own subagents without their own setting, set the CLAUDE_CODE_SUBAGENT_MODEL environment variable. On its own, it doesn't change the built-in Explore and Plan subagents. Anthropic itself notes that subagents can account for a large part of what a session costs, and that early testers had good results with Fable as the planner and other models doing the work.
- Drop “answer directly”, lower the effort. Fable 5.1 usually followed such lines; Fable 5.2 still thinks at medium and above. Only a lower level gets you faster answers.
- Watch for early stops on long tasks. Fable 5.2 sometimes ends a turn before the work is done: it offers to continue, stops at an ordinary obstacle or lists what's left, mostly at low and medium. Anthropic provides a system prompt paragraph for this; in Anthropic's testing, about half as many sessions ended unfinished, with about 30 percent more output tokens. Endings that only list what remains stayed even with it.
- Describe the outcome, not the steps. The Claude Code docs recommend describing the result you want rather than every step, skipping reminders to test because Fable checks its own work more often, and handing over larger tasks in one piece.
Clean up old rules: what Anthropic recommends for Fable 5.2
This is the point that matters most if you run your own rules: Fable 5.2 follows instructions very closely, old ones included. Cautious rules pile up over time in memory, notes and instruction files such as CLAUDE.md, and according to Anthropic the model may have written some of them itself. Fable 5.2 then asks for approvals it doesn't need or holds back routine work. Anthropic saw this in early testing mostly where several agents work together. If you run AI employees with their own rules, skills and notes, clean up before you switch. Here is what the guide recommends; the before/after examples are mine, except in point 2:
1. Delete notes that have lost their reason. The model keeps following an old line like “never retry” long after the reason is gone. Read through memory, notes and instruction files once and remove what no longer has a reason. If an agent is already stuck, Anthropic suggests restarting it with its memory and notes trimmed.
- Before: “Never retry when the tool doesn't respond.”
- After: “If the tool doesn't respond, wait briefly and try again. After the third failure, check in with me.”
2. Turn “never” into a rule with an exception. Fable 5.2 treats “never” as absolute. If you want the model to stay careful with an action rather than avoid it entirely, say so. This example comes straight from the guide:
- Before: “Never delete files.”
- After: “Ask for approval before deleting files.”
3. Write approval rules precisely. Fable 5.2 takes approval rules literally. In one of Anthropic's simulated tests, the system prompt said that “sending messages to people outside this conversation” needed approval, and nobody was there to give it. The model then held back about half of its routine requests to colleagues and other teams. So name exactly the actions that need approval, and list the routine steps the model may take on its own. A general “act more independently” is too vague, according to Anthropic.
- Before: “Get approval before you send any messages.”
- After: “Emails to clients and posts on public channels need my approval. You can write internal notes to other employees and entries on the task board without asking.”
4. Check your stop conditions. A line like “stop and have it reviewed at the first error” can block autonomous work. Remove or narrow such conditions if you no longer need them, for example to errors that change data.
5. Loosen hard style rules. Fable 5.2 follows style rules very closely too. According to the guide, a hard limit such as “one instruction per sentence, in 20 words or fewer” can make its prose choppy. Remove style rules you no longer need and describe the reader instead.
- Before: “Every sentence 20 words at most.”
- After: “Write for readers without specialist knowledge, in full, clear sentences.”
6. Be specific when you approve in the chat. Instead of “continue” or a thumbs-up, say what should happen, for example “Yes, run the migration on the staging database”.
7. Review instructions to delegate. If your skills, agent files or system prompt tell the model to hand work to subagents, and Fable 5.2 starts more subagents than needed in testing, review those lines first.
Anthropic names two limits explicitly. For risky or irreversible steps, keep your own confirmation step, even if you loosen other rules. And changed instructions belong in a new conversation; if you resume a stored API conversation after changing the system prompt, the request can fail. For the switch, Anthropic recommends reading and cleaning up every prompt, skill, agent file, memory file and note you carry over, then running your longest tasks and measuring each change on your own tasks one at a time, including its token cost.
How to use Fable 5.2 in claude.ai and Claude Code
In claude.ai (web, desktop, mobile), pick “Fable 5.2” in the model menu next to the send button. For Cowork, the help center says you need the latest version of Claude Desktop. The same menu has an “Effort” setting that controls how much Fable thinks; in the Claude apps, Fable 5.2 starts at medium.
In Claude Code, switch with /model fable or launch with claude --model fable. From v2.1.321, the fable alias points to Fable 5.2 (in Claude apps gateway sessions, use /model claude-fable-5-2 instead). If Fable bills to usage credits on your plan, the model picker shows “Requires usage credits”. Set the effort with /effort; Fable 5.2 starts at medium here too.
Fable has safety classifiers for cybersecurity and biology. When one fires, Claude Code reruns the request on another model: Opus 4.8 for cybersecurity, Opus 5 for biology. Anthropic says early testing showed about 55 percent fewer blocks than with Fable 5.1.
Data retention: 30 days
For companies, this is the key point: Anthropic's docs say Fable 5.2 on the API requires 30-day retention of inputs and outputs. Without that retention (zero data retention), the model is only available with Anthropic's express authorization. According to the announcement, Anthropic uses it to spot misuse across requests. Eligible enterprise customers can avoid the retention through Enterprise Frontier Safeguards, currently in private beta. The rule is not new: it already applied to Fable 5 and 5.1. If your organization runs under zero data retention, the Claude Code docs say you cannot use Fable there either. What else applies to Claude under GDPR is covered in Claude Privacy and GDPR: What Applies. For your specific setup, we can look at it together.
What is Claude Mythos 5.2?
Mythos 5.2 launched the same day at the same price. According to Anthropic, it is the same underlying model as Fable 5.2, just without certain cybersecurity and biology safeguards. That is why Mythos 5.2 is only available to vetted organizations through the Cyber Verification Program and the Life Sciences Verification Program. Claude Enterprise customers can use Anthropic's security product Claude Security without joining the cyber program, according to the announcement. Outside security research and the life sciences, you will not pick Mythos yourself.
Frequently asked questions
How much does Claude Fable 5.2 cost?
On the API, $8 per million input tokens and $40 per million output tokens, double that in fast mode. On a subscription you don't pay per token: on Max and Ultra, Fable is included up to half of your weekly limit; on Pro it runs on usage credits at API rates outside the launch offer.
Is Claude Fable included in the Pro plan?
Only in the launch week from November 9 to 16, 2026 (ending at 11:59 pm Pacific time): up to half of your weekly limit according to Anthropic, plus one Fable reset. Outside that week, Fable on Pro runs on usage credits. The Free plan has no Fable models.
How much Fable do Max and Ultra include?
Up to 50 percent of your weekly limit at no extra cost. It's a share, not an extra: all models together never go past your weekly limit, and Fable uses it up faster than other models.
What does a Fable limit reset do?
It gives back the Fable usage of your week and refills your five-hour window if you have one. Usage from other models stays where it is, and your limit doesn't grow. You redeem it under Settings > Usage in the browser, the desktop app or the mobile apps, not in Claude Code in your terminal.
What is the difference between Fable 5.1 and Fable 5.2?
On the API, Fable 5.2 costs $8 and $40 instead of $10 and $50 per million tokens, is faster according to Anthropic, and leads in most of Anthropic's own benchmarks. According to Anthropic it follows instructions very closely, old ones included, so it pays to clean up your own rules. On a subscription, both are billed the same way.
How do I switch to Fable 5.2 in Claude Code?
Run /model fable in a session or start with claude --model fable. You need Claude Code v2.1.321 or later; claude update gets you there.
What to do next
Keep Opus 5.5 as your default and use Fable 5.2 where Opus falls short even at higher effort. Before you do, check under Settings > Usage whether your plan bills Fable from your limits or from usage credits, and clean up your AI employees' rules before you move them to Fable 5.2. Which tasks other business owners hand to Fable and when they stay on Opus is what we discuss in my community AI Employee Framework.
