Claude Fable 5 Is in Paid Claude Plans Until June 22: What to Test Before It Switches to Credits
Claude Fable 5 is in Pro, Max, Team, and seat-based Enterprise plans until June 22. What changes on June 23, what the API costs, and what to test this week.
If you pay for Claude, you have a new model available this week: Claude Fable 5, released on June 9. You probably have two immediate questions — do I already have it, and is it going to cost me extra? On Pro, Max, Team, and seat-based Enterprise plans the answers are yes and not yet: Fable 5 is included at no extra cost, but only through June 22, 2026. From June 23, using it on those plans requires usage credits, which Anthropic has not detailed beyond the name. That leaves about ten days to find out whether it earns a place in your daily work without changing what you pay. Here is how to use the window, a workable decision rule for when June 23 arrives, and the two catches worth knowing before you move real work onto it.
What Anthropic shipped, in plain terms
Fable 5 is what Anthropic calls a Mythos-class model, and the version it says it has "made safe for general use." Its sibling, Claude Mythos 5 — the same underlying model with safeguards lifted in some areas — remains limited to vetted cybersecurity partners and selected biology researchers through Anthropic's trusted-access programs. Fable 5 wraps that model in safeguards that screen certain categories of request, and, per TechCrunch's launch coverage, an external bug bounty produced no universal jailbreaks in over 1,000 hours of testing against them.
On capability, Anthropic claims state-of-the-art results on nearly every benchmark it tested — the top score on Cognition's FrontierCode coding evaluation, the highest score of any model on Hebbia's Finance Benchmark, and state-of-the-art results on vision tasks. In the launch discussion on Hacker News — over 2,500 points and 2,100 comments within days — many commenters reported a capability jump over Claude Opus 4.8, including hard refactors finishing in fewer turns. Treat those reports as early anecdotes rather than benchmarks; the volume of attention, at least, says this launch is not being received as a routine version bump.
Which plans include it, and for how long
Your access depends on how you reach Claude. As of June 12, 2026, it breaks down like this:
| Where you use Claude | Through June 22 | From June 23 |
|---|---|---|
| Pro and Max plans | Included at no extra cost | Requires usage credits |
| Team and seat-based Enterprise | Included at no extra cost | Requires usage credits |
| Claude API and consumption-based Enterprise | Pay per token, available now | Unchanged |
Anthropic told TechCrunch it plans to restore Fable 5 as a standard plan feature "as soon as possible," but it has not named a date. Watch for that announcement before you commit to a credit budget; the June 23 change may be temporary, and how temporary is the one thing nobody outside Anthropic knows.
Why some of your answers will come back from Opus 4.8
This is the first catch, and it confuses people mid-session. Fable 5 runs behind classifiers — automated filters that scan each request for cybersecurity, biology and chemistry, and model-distillation topics. When a classifier triggers, your request is answered by Claude Opus 4.8 instead, inside the same conversation. Anthropic says more than 95% of Fable sessions involve no fallback at all, and that the filters are deliberately tuned cautious, so benign requests sometimes trip them.
The launch thread includes several technical false-positive reports. Commenters described fallbacks on MRI brain segmentation code, research about how mosquitoes transmit malaria, and audio firmware development — none of which reads as dangerous. For a developer in security tooling or computational biology, the real risk is subtle: the model you are evaluating may quietly not be the model answering you. Verify which model produced a response before you judge Fable 5 on it — launch-thread commenters could see sessions switching, but Anthropic's announcement does not document how each surface labels a fallback, so find the model indicator in the client you use, if it exposes one. And if your prompts routinely trip those classifiers, be honest with yourself about what your June test measured: a meaningful share of it may have been Opus 4.8 rather than the model you meant to evaluate.
The retention rule your company may care about more than the price
According to TechCrunch, all Fable 5 traffic is retained for 30 days for safety monitoring — including traffic from organizations that negotiated zero-data-retention agreements with Anthropic. Anthropic's stated reason is defending against "complex and novel attacks" on the model's safeguards.
Weigh that against what you send. An individual subscriber drafting emails may shrug; one pasting in medical records or a client's legal documents may not want a 30-day server-side copy, however good the stated reason. And for anyone using Claude under a company agreement, this is the part of the launch that matters most: if your employer chose Anthropic partly because prompts were never stored, that assumption does not hold for this one model. In practice, the move is simple — before routing proprietary code or client data through Fable 5, ask whoever owns your Anthropic contract whether the 30-day rule conflicts with it. That conversation costs a day. Discovering the conflict in an audit costs more.
What it costs once the window closes
On the API, Fable 5 is $10 per million input tokens and $50 per million output tokens — exactly double Opus 4.8's $5 and $25, and the same price Anthropic charges for the fast mode of Opus 4.8, currently a research preview. Fable 5's full 1M-token context window comes at standard rates, and the Batch API halves those rates to $5 and $25 — the same 50% batch discount Opus 4.8 gets.
Double the per-token price does not automatically mean double the spend. If the fewer-turns reports hold up, some tasks may cost less end to end on Fable 5 than on a cheaper model that needs more attempts. That is exactly the kind of claim you should test against your own workload rather than accept, though — and if your team already meters AI spending the way the enterprise AI cost-control playbooks recommend, Fable 5 is a textbook case for measuring cost per finished task, not cost per token.
A sensible way to spend the days before June 22
You have until June 22 to answer one question with your own work: does Fable 5 finish things Opus 4.8 could not, or finish them meaningfully faster? A plan that fits in spare moments:
- Collect the two or three recurring tasks where Opus 4.8 disappoints you — the refactor it keeps fumbling, the analysis it loses the thread on — and run them on Fable 5 this week.
- Before you score any result, check which model actually produced it — mid-session fallbacks to Opus 4.8 are easy to miss. If you cannot tell, leave that response out of the comparison.
- If your work centers on security, biology, or chemistry, expect fallbacks more often and weigh the test accordingly.
- Keep the tasks identical between models where you can; a fair comparison needs the same prompt, not a remembered impression.
- Before June 23 arrives, decide: keep using it on credits, or wait for it to return to plans.
That last decision is easier than it looks. If Fable 5 clearly saved you working hours on tasks you actually repeat, paying for continued access is probably worth it — on the API that price is known (double Opus 4.8 per token), while a subscription credit budget is something you will have to size against your own usage once the credit requirement starts. If your test shows a difference you can only see in benchmarks, stay on Opus 4.8 and lose nothing. What you should not do is decide from launch-day excitement in either direction — the window exists so you can find out for yourself, on tasks that are yours.