Claude Fable 5.2: What CIOs Need to Decide Before Turning It On

Anthropic released Claude Fable 5.2 on 9 November 2026, according to its announcement. AWS and GitHub posted their own availability notices the same day. The conditions attached to the model differ from one channel to the next, and in most of them it cannot run under zero data retention unless an exception applies.

This article sets out what Anthropic and each provider publish as of 11 November 2026, followed by five decisions to make before anyone enables the model. Vendor pages change often, so recheck each linked source before a contract or a rollout depends on it.

What shipped on 9 November

Anthropic shipped one underlying model in two configurations. The system card dated 9 November 2026, which is the vendor's own test report, calls Fable 5.2 and Mythos 5.2 "two configurations of a new large language model". Fable 5.2 is the one on general sale, and it carries added safeguards that stop certain tasks in biology and cybersecurity. Mythos 5.2 has more permissive safeguards, and the card says direct access to it is limited to vetted individuals and organizations.

The published numbers are on Anthropic's model page, read on 11 November 2026. Models are billed per token, a unit of text that the page puts at roughly half a word. The list price is $8 per million tokens sent to the model and $40 per million tokens it writes back. The context window, meaning how much text one request can hold, is 1 million tokens.

Two points are easy to get wrong. The first is the effort setting, which controls how much reasoning the model spends on a request. Anthropic's model page gives the default on its API as high, while the announcement says medium is the default in the Claude apps and Claude Code. The second is a safety level. We found no AI Safety Level label in the announcement or in the text of the system card, which reports against named risk thresholds instead. A risk committee that asks for the level should be told that none is published.

Where it is available, and where it is not listed

The table records what each provider's own page said when we opened it on 11 November 2026. The source for each row is linked in its middle column.

Anthropic's Foundry documentation lists Fable 5.2 only under the option hosted by Anthropic. Microsoft's model list describes that option as running on Anthropic's infrastructure "outside of Azure". Microsoft's page also labels the model Preview and says previews come without a service-level agreement. A team that chose Azure to keep processing inside Azure should read both pages before it deploys.

One place where the model is not listed is Microsoft 365 Copilot. The newest Fable entry on Microsoft's Copilot Cowork "What's new" page, dated 6 November 2026, is Fable 5.1 from September, and the Cowork model list shows the same version. Both pages were last edited before the release. As of 11 November 2026 we found no Microsoft page that says Fable 5.2 is in Microsoft 365 Copilot or gives a date for it, so treat that as unknown.

The 30-day retention requirement

Some enterprises buy AI services on a zero data retention basis, meaning the vendor agrees not to store prompts or answers once the response is sent. Fable 5.2 does not work that way by default. Anthropic's data retention documentation says the model requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it. Anthropic's support article on these models, read on 11 November 2026, says prompts and outputs are kept for 30 days and then deleted unless they were flagged or must be kept by law. It adds that human review happens only through "a controlled access path".

A team that relies on zero retention has published routes forward. Anthropic's data retention documentation says an organization can turn on 30-day retention for one workspace while its other workspaces keep zero retention. On Foundry the setting applies to a whole Azure subscription, and Microsoft's deployment guide says a subscription with zero retention receives a 400 error until that is changed with Anthropic or a new subscription is used. The alternative is a model without the requirement. The GitHub changelog says Claude models other than the Fable versions continue under zero data retention.

There is also a temporary exemption, and the providers describe its length differently. Anthropic's product page says it is building Enterprise Frontier Safeguards, which runs the safety monitoring inside the customer's own infrastructure. Until that is available, the page says, eligible customers can keep using Fable 5.2 with zero retention. The GitHub changelog says eligible enterprises have the exemption "through the end of the calendar year" and will need the new safeguards after that. The AWS post says zero retention for eligible customers is available for internal use "through at least June 30, 2027". None of these pages publishes the criteria for eligibility.

What the system card says about long tasks

The main third-party result on long tasks in the system card is FrontierSWE v2, a benchmark of 34 engineering and research tasks with a 20-hour budget for each attempt. Its maker, Proximal, ran the test, and Fable 5.2 scored 69.9% against 62.3% for Opus 5.5. That is a result on set tasks, and it says little about your own systems.

The same system card shows where the model does not lead. On GDPval-AA, a knowledge work benchmark from Artificial Analysis where a higher rating is better, Fable 5.2 was rated 1831 against 1866 for Opus 5.5. On the FrontierCode 2.0 coding benchmark, the headline score of 49.1% came at the second-highest effort setting. At maximum effort the score was 45.7%, and the card says that counting 22 timed-out attempts as zero would put it at 44.4%, below Opus 5.5 at 45.5%.

When a safety classifier blocks a cybersecurity request, Anthropic's own products pass it to an older model, Claude Opus 4.8. The system card reports a prompt injection test, where hidden instructions in content try to hijack an agent. In it, 53% of attempts against Fable 5.2 fell back to Opus 4.8. The chance that an attacker succeeded within 15 attempts was 4.09%, against 0.96% for Opus 5.5. That is a test result and should not be read as a rate seen in production. The card says fallback may differ on other platforms, and the AWS post says that on Bedrock "the retry is your code's job".

The decisions to make before you turn it on

Which workloads justify it

The announcement itself recommends starting with Claude Opus 5.5 for most coding and knowledge tasks and moving up only for harder work. Fable 5.2 costs twice as much per token on Anthropic's price list, at $8 and $40 against $4 and $20 for Opus 5.5. A sensible first list holds the long engineering or analysis tasks where your current model gives up or needs frequent correction.

Which platform

Choose the channel by where your data is allowed to go. The AWS post says Amazon personnel perform the safety review inside the AWS boundary, but Bedrock offers no Europe-only routing today. Google Cloud offers Europe routing and asks for prompt-response sharing with Anthropic. On Foundry the bill comes through Azure while requests run on Anthropic's infrastructure. Routing geography does not settle every data residency question, so record which channel your data classification policy permits.

Who signs off on retention

Retention is a decision for the people who own privacy and client contracts. Ask them whether any client agreement or internal policy promises that prompts are never stored by a model vendor. If one does, the per-workspace and per-subscription settings described above let you keep Fable 5.2 away from that data. If you are offered the temporary exemption, get its end date in writing from the provider you buy through, because GitHub and AWS publish different dates. This article is general information and not legal advice.

How you will test it on your own tasks

Run the model on a fixed set of your own tasks, and score the results the same way you score your current model. Set the effort level explicitly in every test, since the default is high on the API and medium in the Claude apps. Include tasks that touch security topics so you see how often requests are refused or handed to another model. Our guide to regression testing when models change describes the method.

What a finished task costs

The price per token says little about the bill, because effort changes how many tokens a task consumes. The system card reports that on FrontierCode 2.0 the average cost per attempted task for Fable 5.2 rose from $2.10 at low effort to $44.70 at maximum effort. At high effort it scored 45.3% for $6.72 a task, while Opus 5.5 reached its top score of 45.5% at $12.82 a task. The model with the higher token price was cheaper per task at one setting and several times more expensive at another. Divide your own cost by the share of tasks that pass review, as set out in our article on budgeting agents by cost per finished task.

Where QueryNow fits

QueryNow builds the foundation enterprise AI runs on, and the agents that run on it. Matching a model to each workload and checking each cloud's data terms are part of that work on Azure and AWS. If you are weighing Fable 5.2 against the model you use today, you can talk to our team about the workload you have in mind.