Anthropic, the company behind the AI assistant Claude, released Claude Fable 5.2 and Claude Mythos 5.2, new versions of its most powerful models, on November 9, 2026. Only Fable is open to every paying customer. For developers who pay per use, Fable 5.2 costs 20% less than Fable 5.1. Subscribers to Pro, Claude’s entry-level paid plan, can also use it at no extra cost for one week.
Prices on the API, the pay-as-you-go access developers use to build Claude into their software, are set per million tokens. Tokens are the word fragments a model reads (input) and writes (output). Fable 5.2 drops from $10 to $8 for input and from $50 to $40 for output.
Fable sits at the top of the Claude lineup, above Opus, Sonnet and Haiku (our guide to Claude models). Yet Anthropic recommends starting most coding and knowledge work on Claude Opus 5.5, which costs half as much on the API.
One model with two sets of safeguards
Fable 5.2 and Mythos 5.2 run on the same underlying model, according to Anthropic’s announcement. Fable has extra restrictions in cybersecurity and biology to prevent misuse, which lets Anthropic offer it to every paying customer. Mythos is reserved for organizations accepted into one of two vetting programs, one for cyber defense and one for biomedical research. Claude Enterprise customers can also use Mythos through Claude Security, a tool that scans code for vulnerabilities.
On Fable, automated filters block requests that Anthropic considers high risk. Examples include penetration testing, meaning simulated attacks that look for a system’s weak spots, and writing exploits, programs that take advantage of a security flaw. The filters also block most requests in research biology.
When a filter steps in, a less capable model answers instead: Opus 4.8 for cybersecurity, Opus 5 for biology. Anthropic says early testing across its apps showed about 55% fewer blocks than with Fable 5.1.
The filters also read files, web search results and content from connectors, the outside services linked to Claude, according to a separate Anthropic help page. So a block can come from text the user never typed.
Fable 5.2 prices and specs
Fast mode, an option that generates output up to 2.5 times faster, doubles Fable 5.2’s price to $16 for input and $80 for output. Opus 5.5 costs $4 for input and $20 for output. In euros, Fable 5.2’s standard price comes to about €7.15 for input and €35.74 for output. The conversion uses the European Central Bank’s reference rate for November 9, 2026: $1.1193 to the euro.
Fable 5.2 accepts up to one million tokens of context, the amount of text it can work with at once. It writes at most 128,000 tokens per response, and Anthropic says it produces output up to 30% faster than Fable 5.1. It is available on Anthropic’s own platform and through the cloud providers Amazon Web Services, Google Cloud and Microsoft Azure.
Fortune framed the release as Anthropic’s answer to GPT-6 Astra, the OpenAI flagship released in early September. According to Fortune, Astra costs $10 for input and $50 for output per million tokens, the old Fable 5.1 price.
The launch offers, plan by plan
Pro subscribers get Fable for one week, according to Anthropic’s help page on the launch offers. So do standard seats, the basic licenses on Claude’s Team plan. The offer runs from November 9 through 11:59 p.m. Pacific on November 16. During that week, Fable can account for up to 50% of the weekly usage limit.
Pro and standard Team accounts also get one reset, usable until the same deadline, which wipes out the usage already spent on Fable. After the week, these users pay for Fable with usage credits, billed on top of the subscription at API rates.
The Max and Ultra plans, Claude’s higher-priced tiers (our story on the Ultra plan), already include Fable for up to half of the weekly limit. So do premium seats, the upper tier of business licenses. These accounts get one to three resets, depending on how long they have been subscribed: one for under six months, two for six to 12 months, three for more than 12 months. Those resets expire at 11:59 p.m. Pacific on March 8, 2027.
In both cases, resets go only to accounts that had paid billing in place before 10 a.m. Pacific on November 9. The free plan and several versions of the Enterprise plan are excluded.
On Enterprise plans, Fable 5.2 stays off until an administrator turns it on. Seven days after launch, it will be added to every role that already has access to a Fable model, unless an administrator opts out first. A role is a group of user permissions.
A reset erases only the weekly usage spent on Fable and does not raise any limit. It also fully refills the five-hour usage window on accounts that have one. In Anthropic’s example, a subscriber has used 80% of the weekly limit: 50% on Fable, 30% on other models. After a reset, that subscriber is back at 30%.
Fable models use up limits faster than other models, the help page warns. On the evening of November 9, several Reddit users asked whether a reset gives back the whole week’s limit. One said a single Fable task used up 15% of the weekly limit. Anthropic already faces a class action in the United States over the way it describes its usage limits (our story on the case).
Anthropic’s own test results
The first test in Anthropic’s comparison table, FrontierCode 2.0, checks whether the maintainers of open-source projects would accept a model’s code changes. Scores depend on effort, a setting that controls how deeply the model reasons. At very high effort, Fable 5.2 scores 49.1%, ahead of Opus 5.5 at 45.5% and GPT-6 Astra at 42.8%. Google’s Gemini 4 Carbon scores 41.1%, measured at medium effort (our story on Gemini 4 Carbon).
At plain high effort, Fable 5.2 reaches 45.3%, close to Opus 5.5’s best score. Anthropic says it gets there for about half of Opus 5.5’s cost per task.
Anthropic also put about 1,000 Fable 5.2 agents, copies of the model that split up the work, on the proof of Fermat’s Last Theorem. In 30 hours, they translated it into a language a computer can check, the company says. Its first such translation, by an unreleased research version of Claude, took 11 days. The company also says Fable 5.2 solved two open problems in black hole physics and designed three electronic devices that worked without electrical modifications.
Caveats and open questions
These scores and demonstrations come from Anthropic’s own tests or from testers the company picked. Outside evaluators, including Britain’s government-run AI Security Institute, tested the model before release. But a note in the announcement says they did not review the announcement or the model’s technical report.
CNBC noted that this is Anthropic’s first release of frontier models, the industry’s most advanced systems, since September. That month, its CEO, Dario Amodei, called for slowing the pace of AI development. “We called for pacing, not pausing,” Dianne Penn, Anthropic’s head of product management for research, told CNBC.
Anthropic also reports that Mythos 5.2 and Fable 5.2 say they suspect they are being evaluated. They say so more often than any of the company’s recent models, “which challenges our ability to assess how they will act in real-world settings.” TechCrunch also quoted the technical report: on several evaluations, Mythos 5.2 is better than earlier Claude models at acting without being detected. Anthropic says Mythos 5.2 still cannot act covertly against its user.
Another test, with more than 1,800 scenarios, tried to tempt the model out of the isolated environment it runs in during testing. Mythos 5.2 used credentials or files placed outside that environment in 3.1% of cases, close to Opus 5.5 at 3.3%. Earlier Mythos versions did it much more often: 21% for Mythos 5.1, 39% for Mythos 5. An incident involving Mythos 5 during a hacking test became public in August (our story on Mythos 5).
According to the API documentation, Fable 5.2 requires 30-day data retention for safety monitoring. Only customers Anthropic authorizes can skip it, for example through Enterprise Frontier Safeguards. That setup, which monitors data stored in the customer’s own systems, is in private testing with Bridgewater, Salesforce, Stripe and other customers.
On Reddit, one user said people they know outside the United States see every request switched to Opus. Anthropic’s help pages tie these switches to the topic of a request and mention no rule based on location. We could not verify the report.
What it means
Fable 5.2 is cheaper than Fable 5.1 but still costs twice as much as Opus 5.5, the model Anthropic recommends by default. For Pro subscribers, the launch offer amounts to a one-week trial. On Max and Ultra, resets only clear the Fable share of the weekly limit.
The price cut can be checked on Anthropic’s public price list. On performance, none of the sources cited here offers an independent measurement. The 30-day retention rule could be a hurdle for organizations covered by the GDPR, the European Union’s data protection law, if they send sensitive documents to the API.
The Pro week ends at 11:59 p.m. Pacific on November 16. For Enterprise customers, the automatic rollout comes seven days after launch. Next to watch are the first independent performance rankings, how much of their limits subscribers actually spend on Fable, and Anthropic’s stock market listing. According to CNBC, the listing could come as soon as next week.
Sources
- Introducing Claude Fable 5.2 and Claude Mythos 5.2, Anthropic, November 9, 2026
- Claude Fable 5.2 launch offers, Claude Help Center, accessed November 10, 2026
- Claude Fable models on your plan, Claude Help Center, accessed November 10, 2026
- Pricing, Claude Platform documentation, accessed November 10, 2026
- Claude Fable 5.2, Claude Platform documentation, accessed November 10, 2026
- Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 ahead of IPO, CNBC, November 9, 2026
- Anthropic debuts updated Claude Fable and Mythos models, priced 20% below OpenAI’s GPT-6 Astra, Fortune, November 9, 2026
- Anthropic releases new Mythos and Fable models amid RSI concerns, TechCrunch, November 9, 2026
Header image: MACCUS chart, data from Anthropic (Claude API pricing page, accessed November 10, 2026)
