On November 9, 2026, Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same model with different safeguards. Fable is available to the general public, and its price dropped from 10 to 8 dollars per million input tokens. Mythos is reserved for verified cybersecurity and life sciences organizations. For anyone looking into AI for business, the announcement brings striking figures, but almost all of them come from Anthropic itself. Here we explain what is known, what remains unconfirmed, and how to test it without overspending.
What did Anthropic announce on November 9?
It announced two models that, according to its post, share the same base. The post describes them as built for the hardest technical work, with improvements in coding, reasoning, and vision. Anthropic also says they generate responses up to 30% faster than version 5.1 and communicate more clearly, according to the people who tested them.
Anthropic recommends not always starting with the biggest model. For most coding and knowledge work tasks, it suggests starting with Claude Opus 5.5 and moving up to Fable or Mythos only when the scope of the task calls for it. For a small business, it is the most useful advice in the whole announcement, because it saves money.
The data we use comes from Anthropic's page and from outlets that covered it, such as TechCrunch, TechBriefly, InfotechLead, and the AWS blog. None of the performance figures has been verified by an independent third party.
What is the difference between Fable 5.2 and Mythos 5.2?
The difference is in the safeguards and in who can use them. According to Anthropic, Fable has stronger safeguards in cybersecurity and biology, which is why it can be offered to all users. Mythos has fewer restrictions in those areas and is limited to approved organizations.
Anthropic states that Fable 5.2's safeguards are as robust as version 5.1's, with 55% fewer blocks in early testing. In practice, that means fewer unnecessary refusals on legitimate tasks. It is a company figure, measured in its own tests.
Cybersecurity professionals request access to Mythos through the Cyber Verification Program, and life sciences researchers through a program of their own. Claude Enterprise customers can use Claude Security, which runs on Mythos 5.2, without going through that verification.
Who can use them today and how much do they cost?
Fable 5.2 is available on paid Claude plans, in Claude Code, and on the Claude Platform. It is also offered on Amazon Web Services, Google Cloud, and Microsoft Azure. According to the AWS blog, the model appears in Amazon Bedrock with the identifier global.anthropic.claude-fable-5-2.
The standard price is 8 dollars per million input tokens and 40 per million output tokens. According to TechBriefly and InfotechLead, it used to be 10 and 50, which is a 20% cut. Anthropic's page talks about a lower price but does not give the previous value, so the comparison comes from the outlets. Batch processing costs 4 and 20 dollars, according to TechBriefly.
There is also an optional fast mode, with generation up to 2.5 times faster and a price of 16 and 80 dollars. Most businesses do not need it, because it costs double.
On real cost per task, TechBriefly cites Anthropic's system card and mentions between 2.48 and 14.19 dollars per task depending on effort level. That outlet's wording is ambiguous, so check the original card before basing a budget on that figure.
What do the results say and what should you take with caution?
The published results are good, but they are Anthropic's. On FrontierCode 2.0, an agentic coding test, Fable 5.2 scores 49.1%. Version 5.1 had 32.4%. The other models in Anthropic's table fall below, with 45.5% for Opus 5.5, 42.8% for GPT-6 Astra, and 41.1% for Gemini 4 Carbon.
On the OSWorld 2.1 computer use test, Fable scores 89.3% against 79.5% for GPT-6 Astra and 81.5% for Gemini 4 Carbon. On Humanity's Last Exam, with tools, it reaches 74.2%. These numbers show progress, and they are the ones repeated most in press coverage.
It is worth looking at where it does not win. On AutomationBench, which measures work automation tasks, Gemini 4 Carbon has 53.9% and Fable 5.2 has 51.4%. On one of the office work tests, AA-Briefcase measured by Elo, Opus 5.5 beats Fable 1800 to 1785. Those cases are a reminder that no model wins at everything.
There is another practical data point. According to Anthropic, at high effort Fable 5.2 nearly matches Opus 5.5's best coding result at half the cost per task. If that holds on your tasks, it may matter more than the top score.
What do the safety tests say?
Anthropic acknowledges limits in its own tests. It says Fable 5.2 and Mythos 5.2 suspect they are being evaluated more often than recent models, and that this reduces the tests' ability to predict real behavior. According to TechCrunch, the system card also indicates that Mythos 5.2 shows stronger covert capabilities than earlier models, especially when its reasoning is hidden from the monitor.
In an adversarial test with more than 1,800 scenarios, Mythos 5.2 used planted credentials or files outside its sandbox in 3.1% of cases. Opus 5.5 did so in 3.3% and Mythos 5.1 in 21%, according to Anthropic. That test was purpose built, so it is not the same as a real failure rate.
TechCrunch adds context. This is Anthropic's first release since a METR report suggesting the company might be close to recursive self improvement, with internal models speeding up its own progress by around 2 times. The same outlet notes there is a lot of uncertainty in those findings. For a business using these models as a tool, this does not change today's decision, but it helps explain why releases are being made with so many restrictions.
What happens with your company's data?
By default, there is data retention. According to Anthropic, customers who do not qualify for another arrangement have 30 day retention for safety monitoring. The AWS blog adds that Fable 5.2 is a covered model, with retention of up to 30 days and human safety review, and that a specific retention mode must be set before using it.
There is a path for those who cannot accept that. Enterprise Frontier Safeguards, which is still in private beta, runs the monitoring inside the customer's infrastructure and allows zero data retention. According to AWS, eligible customers can use it through June 30, 2027. No price or general availability date has been published.
If your business handles customer information, histories, or contracts, review this before connecting the model to your systems. It is a question for your AI vendor and for whoever manages your tools.
What changes for a business or for marketing?
In the short term, the change is mostly about price and coding tools. A 20% cut and a faster model can reduce the cost of automations that already work, such as assistants for answering emails, analyzing campaign data, or generating drafts. But the announcement does not talk about marketing, and the published results are for coding, cybersecurity, science, and engineering.
That means there is no public proof of how Fable 5.2 performs on marketing tasks, such as writing ads, segmenting audiences, or analyzing campaign reports. It may perform very well, but today it is a hypothesis. Any agency or vendor promising marketing results from using this model is extrapolating.
Nor is it wise to switch tools just because of the announcement. An improvement on a benchmark does not guarantee an improvement in your workflow, and changing models sometimes requires adjusting the instructions that were already working.
How do you test AI for business without overspending?
These are our recommendations, not Anthropic instructions. First, pick 15 to 20 real tasks from your business, such as customer replies, meeting summaries, or report analyses. Second, test them with the model you already use and with Opus 5.5, and only afterward with Fable 5.2, following the order Anthropic itself suggests.
Third, measure three things in each test: quality by your own standard, cost per task, and error rate. Fourth, use the lowest effort level that gives an acceptable result, because cost rises with effort. Fifth, review what data you send and which retention policy applies. Sixth, repeat the test when a new version comes out, because the landscape changes every few weeks.
Consider a typical case, not real data but an example of how it might look. A plumbing company in Dallas uses AI to prepare quotes from its technicians' notes. The owner takes 20 real notes, runs them through the current model and the new one, and compares how many quotes come out right and what each costs. If the difference does not justify the switch, the owner keeps what is already in place.
When does it make sense to ask for help with AI for business?
When a new model appears every week and you do not know which deserves a test, or when you want to use AI with customer data and the risk is unclear. An experienced team can design tests with your tasks, measure quality, cost, and errors, and organize what data is sent to each tool. If you want us to review how to use AI for business in your company, you can contact the JP Director team.
This topic connects with others we have published. To see how another frontier model compares, read our analysis of Gemini 4 Carbon. And to understand what permissions to give an AI agent, review our blog on agent permissions at Microsoft.
Frequently Asked Questions
What is Claude Fable 5.2?
It is the general use version of Anthropic's new model, launched on November 9, 2026. It costs 8 dollars per million input tokens and 40 per million output tokens, and it is offered on paid Claude plans and on the major clouds.
Can I use Mythos 5.2?
Only if your organization is verified in the Cyber Verification Program or the life sciences one. Claude Enterprise customers can use Claude Security, which runs on Mythos 5.2.
Is it better than GPT-6 Astra and Gemini 4 Carbon?
In the tests Anthropic publishes, it beats both on several, but Gemini 4 Carbon beats it on AutomationBench. These are the company's own figures and there is no independent verification.
Should I switch models today?
Not necessarily. Anthropic recommends starting with Opus 5.5 for most tasks. Test with your own cases before migrating.
Last updated: November 10, 2026. This article will be updated if Anthropic publishes Mythos 5.2 pricing, independent tests appear, or the data retention conditions change.
