Claude Fable 5: A New Tier Above Opus, and a New Way of Working

Claude Fable 5 is Anthropic's new top model, a tier above the Opus line, and it arrived in Claude Code priced at $10 per million input tokens and $50 per million output tokens. That is exactly double Opus 4.8. The pitch is not "smarter answers." It is a different working relationship: a model that investigates before it acts, holds a multi-day task without losing the thread, and checks its own work before telling you it is done. Anthropic's phrase is that it works "more like a seasoned engineer," and the launch guidance that comes with it reads less like a feature list and more like instructions for managing a new senior hire.

Three things in this release are worth understanding before the free window closes on June 22: what the model actually changes about day-to-day use, what the pricing implies about where Anthropic thinks the value is, and how the safeguard architecture works, because it is the first time Anthropic has shipped a frontier model whose distribution strategy is built around routing.

What Is Claude Fable 5?

Claude Fable 5 is Anthropic's next generation of intelligence, positioned above Opus 4.8 in the lineup and available now in Claude Code (select it with /model) and through the API as claude-fable-5. It carries a 1 million token context window and supports up to 128K output tokens, matching the recent Opus models on raw capacity. The API surface is essentially the same as Opus 4.8: adaptive thinking only, no manual thinking budgets, no sampling parameters. For most integrations, switching is a model ID change.

The differentiation is behavioral. Anthropic describes three habits that define the model: it investigates your codebase before it acts, it pulls in relevant context as it works rather than relying only on what you handed it up front, and it reviews its own output before reporting completion. Each of those habits targets a specific failure mode that anyone who has run agentic coding tools at scale will recognize: the confident edit made without reading the surrounding code, the fix that ignored a convention documented two directories up, and the "done" that was not.

The other emphasis is duration. Fable 5 is built for long-running, asynchronous work, the kind of multi-hour or multi-day session where you hand off a task and walk away. That capability has been creeping into the agentic tooling landscape for a year, but pairing it with a model explicitly trained to verify its own work is what makes the handoff credible. An agent that runs unattended for six hours is only useful if you can trust its definition of finished.

How Much Does Claude Fable 5 Cost?

On the API, Fable 5 costs $10 per million input tokens and $50 per million output tokens. Here is where that sits against the rest of Anthropic's current lineup:

For subscribers, there is a meaningful launch window: through June 22, 2026, Fable 5 is included in Claude subscriptions and draws from your plan's normal usage, with no additional cost. After that, it runs on usage credits. In other words, Anthropic is giving every subscriber roughly two weeks of free benchmarking against their real workload, and that window is the most valuable part of the launch for anyone deciding whether to adopt.

The 2x premium over Opus is a statement about where Anthropic thinks the economics of these models have moved. A per-token price only matters relative to how many tokens a task consumes, and how many attempts a task takes. If Fable 5 completes a gnarly migration in one unsupervised overnight run that Opus would have needed three supervised attempts at, the expensive model is the cheap one. That is the same lesson that surfaced in Uber's AI spending reckoning: tokens are an input, and the only number that matters is cost per delivered outcome. A premium-priced model raises the stakes on measuring that properly, because at $50 per million output tokens, an unmeasured workflow gets expensive fast.

How Does Working With Fable 5 Differ From Previous Models?

The most interesting part of the launch is the usage guidance, because it asks users to unlearn habits that the last two years of AI tooling taught them. Anthropic's recommendations come down to four shifts.

Give it the goal, not the steps. Instead of decomposing a task into a checklist and feeding the model one instruction at a time, describe the outcome and let the model plan the path. In Claude Code this pairs with the /goal command, which keeps the model working until a completion condition is met. This is a real inversion: prompt engineering as most teams practice it is precisely the art of writing better step lists.

Size up asynchronous tasks and walk away. Work you would normally break into supervised chunks can be handed over whole. The model is built to hold multi-hour and multi-day sessions without losing the thread, which changes the unit of delegation from "a prompt" to "a project."

Skip the guardrail prompts. Anthropic says explicitly that you no longer need to write things like "make sure to test" because the model reviews its own work by default, and that qualitative feedback works better than itemized instructions. Most engineering organizations have accumulated prompt libraries full of exactly this boilerplate. If the guidance holds up in practice, those libraries shift from asset to noise, and the differentiating skill becomes writing a good problem statement rather than a good instruction list.

Hand it ambiguous problems. Root-cause analysis, outage debugging, architecture decisions. Anthropic claims Fable 5 more reliably identifies underlying causes instead of patching symptoms. This is the claim with the highest payoff if true, because symptom-patching is the most expensive failure mode of AI-assisted engineering: it produces changes that look like progress while leaving the actual defect in place to resurface later.

Taken together, the guidance describes a management change more than a tooling change. The skills it rewards (writing clear success criteria, scoping a multi-day task, reviewing finished work rather than supervising work in progress) are the skills of delegating to a senior person. Teams that already operate that way will adapt in a day. Teams whose AI usage is built around tight supervision loops will need to actually change how they work to get what they are paying the premium for.

How Do the Safeguards and the Mythos Connection Work?

Fable 5 shares its underlying model with Claude Mythos 5, the frontier system that has mostly existed in public view as a name attached to capability and security speculation. The difference between the two is not intelligence but packaging: Fable 5 ships with additional safeguards, particularly in the cyber and bio domains, and Anthropic states plainly that those safeguards are what allow it to share this level of capability broadly.

The mechanism is notable. When a request touches a high-risk area like cybersecurity or biology, Fable 5 may route it to Claude Opus 4.8 rather than answering with the frontier model. That is a different architecture from the refusal-based safety everyone is used to. Instead of the top model declining, the request silently gets a less capable but well-characterized model. For most users this will be invisible. For teams doing legitimate security work, it is worth knowing about, because it means the model serving your request can vary by topic, and behavior on security-adjacent tasks may not match behavior everywhere else.

There is also a data condition attached: Fable 5 requires a limited retention period, with retained data used only to detect and prevent serious misuse, and not used to train Claude. The framing is straightforward. Frontier capability comes bundled with monitoring sufficient to catch serious abuse, and the no-training commitment is the counterweight that makes the retention palatable. Organizations with strict data-handling requirements should read the retention details before routing sensitive workloads to it, the same diligence any new model tier deserves.

Should You Switch to Claude Fable 5?

The honest answer is: measure it during the free window, against your own work, and let the results decide. Three suggestions for doing that well.

1. Benchmark on the tasks you currently avoid delegating. Comparing Fable 5 to Opus 4.8 on tasks Opus already handles well tells you nothing useful, because Opus is half the price. The relevant test set is the work you have not trusted a model with yet: the multi-day refactor, the intermittent bug nobody has root-caused, the architecture question with no clean answer. That is the work the premium is priced against.

2. Track cost per finished task, not tokens. A more autonomous model consumes more tokens per session by design, since it investigates, verifies, and iterates on its own. Judged on token burn it will look expensive. Judged on completed work per dollar and per hour of human attention, the comparison can flip. Decide which number you are optimizing before the bills arrive, and if usage limits are a concern, our guide on managing Claude token consumption covers the mechanics.

3. Keep the tiers straight. Fable 5 does not replace the rest of the lineup, it extends it upward. Sonnet 4.6 remains the volume workhorse, Opus 4.8 the strong default (we covered what Opus 4.8 changed when it landed), and Fable 5 the tier for the hardest, longest, most ambiguous work. Routing everything to the top model is the same mistake as routing everything to the cheapest one: a failure to match the tool to the task. For subscription pricing across the Claude lineup, see our Claude pricing breakdown.

The larger signal in this launch is the direction of travel. Every release in this cycle has moved the frontier away from "a better autocomplete" and toward "a colleague you brief." Fable 5 is the most explicit statement of that yet, right down to launch guidance that tells you to stop writing instructions and start writing goals. The model is new. The management skills it rewards are not, and the teams that already have them will get the most out of it.

Frequently Asked Questions