Claude Fable 5.2 Requires 30-Day Retention for Most Buyers

Keep Claude Opus 5.5 as your default and send Fable 5.2 only the work Opus fails. Anthropic released Claude Fable 5.2 on November 9 at $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1 and still twice Opus 5.5's $4 and $20. Anthropic's launch post calls Opus 5.5 "our strongest daily driver" and tells users to start there. Most customers who move up also accept 30-day retention of prompts and completions.

Your decision this quarter has three parts: which workloads justify the higher rate, what effort level to run them at, and whether your data can sit in a 30-day safety log. Anthropic's benchmarks answer only the first, and only on its own terms.

What Anthropic Shipped on November 9

Anthropic shipped one underlying model under two names. Fable 5.2 carries cyber and biology safeguards and is open to all customers; Mythos 5.2 lacks certain of those safeguards and goes only to vetted organizations. CNBC reports that Fable 5.2 went generally available the same Monday, and it is listed on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

Rate per million tokens	Fable 5.2	Fable 5.1	Opus 5.5
Input	$8	$10	$4
Output	$40	$50	$20
5-minute cache write	$10	$12.50	$5
Cache read	$0.20	$0.25	$0.20

Prices are from Anthropic's pricing page, checked November 9, 2026. Fast mode doubles Fable 5.2 to $16 and $80. TechCrunch notes the new input price undercuts GPT-6 Astra.

---

Is Fable 5.2 Cheaper Than Opus 5.5 per Task?

Anthropic's published per-task savings for Fable 5.2 all compare it with Opus 5.5 at Max effort or at Opus's top score. Effort is the API control that sets how much thinking, latency and token spend a model puts into a request. Anthropic says Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort on FrontierCode 2.0, 45.3% against 45.5%, for roughly half the cost per task.

If you run Opus 5.5 at Max today, that claim deserves a test. A team that never touched the setting is on medium, the Opus 5.5 API default. The launch post's text gives no dollar costs and no FrontierCode score for Opus 5.5 at Medium, so the half-cost claim can't tell you what happens to a bill built on the default.

The default also moves when you swap model IDs. Anthropic's migration guide warns that a request omitting effort runs one level higher on Fable 5.2, and the change "shows up only as different cost, latency, and results." You get no error to catch it. The same swap on the Claude apps and Claude Code behaves differently, since Medium is the default there.

Cache reads narrow the gap on long agent sessions, because they cost $0.20 per million tokens on both models. Take a session with 2 million cache reads, 100,000 fresh input tokens and 50,000 output tokens, and assume both models use the same token counts. Fable 5.2 costs $3.20 and Opus 5.5 costs $1.80, a 1.78x difference at list price. Output is 63% of the Fable bill in that example, so the number to watch in your own logs is output tokens per accepted result.

Lower effort has a cost of its own. Anthropic documents that on long, multistep tasks Fable 5.2 sometimes ends a turn with a list of what remains, more often at low and medium. Count those early stops as failures in any cost comparison.

Where Fable 5.2 Leads, and Who Checked

Fable 5.2 tops most of Anthropic's published table, and every number in that table comes from Anthropic's launch materials. At Xhigh effort it scored 49.1% on FrontierCode 2.0, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. Cognition, which built FrontierCode, published the same Fable 5.2 and Opus 5.5 scores, so they are not an outside check. The agentic search benchmark WANDR shows 82.1% against 72.3% for Opus 5.5.

It does not win everything. Gemini 4 Carbon leads AutomationBench, a business workflow test, 53.9% to 51.4%. Opus 5.5 keeps the higher AA-Briefcase Elo, 1800 to 1785, on a knowledge work test where Fable 5.2 has the higher rubric score. If your workload looks like document production, that split is worth more attention than the coding charts.

As of November 9, two outside testers have published results, and one disagrees with Anthropic about effort. Artificial Analysis, which runs its evaluations independently, lists Fable 5.2 on FrontierCode 2.0 at 49.1% on High, 47.6% on Xhigh and 48.8% on Max, while Anthropic reports 45.3% on High and 49.1% on Xhigh. Vals AI, testing at High effort, ranks Fable 5.2 first on its Vals Index at 71.31%, ties it with Opus 5.5 on Terminal-Bench 4.0 and puts it behind Opus 5.5 on SWE-bench Multimodal, 62.29% to 70.21%.

Treat the rest as vendor claims for now. Implicator reports that the speed gain, the 55% reduction in blocks and the science results could not be independently verified at publication. Anthropic's own footnote says the outside safety testers, including METR and the UK AI Security Institute, did not review the post or system card and may not agree with its claims. The post also concedes that both models voiced suspicion they were being evaluated more often than recent models, which makes test results harder to read. TechCrunch quotes the system card saying Mythos 5.2 "shows stronger covert capabilities than prior Claude models on several of our evaluations."

The strongest argument for moving is Anthropic's own routing rule, and it is a narrow one. Its documentation says to move to Fable 5.2 when your evals on an Opus model at higher effort still don't meet your requirements.

---

What 30-Day Retention Means for Regulated Data

Unless Anthropic authorizes an exception, Fable 5.2 requires 30-day data retention and is unavailable under zero data retention. Zero data retention, or ZDR, is a contract arrangement under which the provider does not store prompts and outputs after processing them. Anthropic's Covered Models page says prompts and completions are kept for "at least 30 days", then deleted, unless the data is under a safety investigation or must be held for legal reasons.

The failure is loud on the Claude API. If your organization runs Opus 5.5 under ZDR and you add Fable 5.2 to a router, every request returns a 400 invalid_request_error until you change the setting. A ZDR organization can turn on 30-day retention for one named workspace and leave the others at zero, though the Default Workspace always follows the organization's setting. On Bedrock and Google Cloud's Agent Platform, retained data stays inside the cloud provider's environment, which may change the answer your privacy team gives.

Enterprise Frontier Safeguards (EFS) is Anthropic's route around the requirement. EFS runs automated safety monitoring on prompts and responses stored in the customer's own infrastructure, which makes it compatible with ZDR. It is in private beta with customers including Bridgewater, Salesforce and Stripe, and eligible enterprises can use Fable 5.2 with ZDR before they implement it. The Covered Models page limits that interim option to internal business applications, so a customer-facing product needs a separate answer.

EFS moves work to you as well. When monitoring flags a pattern, the signals go to your team, and The Register described this in September as shifting safety work to customers. Someone on your side has to own that queue.

Anthropic has said it introduced retention because correlating misuse across sessions and accounts requires keeping data, and that the policy "was not motivated by a desire to train on enterprise data." Buyers pushed back anyway. Crypto Briefing reports that Microsoft limited internal staff access to Fable 5 on June 10 over customer confidentiality, the day after that model launched, and that finance, law and healthcare customers reacted poorly. Anthropic has acknowledged that many enterprises, especially in regulated industries, found it hard to use models with retention.

What Breaks When You Move Traffic

Several Fable 5.2 differences show up only in production. Rate limits are the first: Fable 5.2's limit is separate from Opus 5.5's, with lower token limits at each tier. In organizations using customer-managed encryption keys, structured outputs are unavailable on Fable 5.2, though Opus 5.5 supports them.

Routing direction matters. On the Claude API and Google-managed Google Cloud, Fable 5.2 can read Opus 5.5's thinking blocks, and Opus 5.5 cannot read Fable 5.2's. Thinking blocks are the stored reasoning a model sends back on later turns. Anthropic says to expect higher cost and latency on the first turn after a switch down, so escalate from Opus to Fable within a conversation and avoid bouncing back.

Older accounts also lose a grace period. Opus 5.5 enforces its check against edited conversation history (the prefix check) by default only on accounts created on or after August 31, 2026, while Fable 5.2 enforces it on every account. Code that rewrites earlier turns can work on Opus today and fail on Fable tomorrow.

Refusals need their own handling. Fable 5.2's classifiers block high-risk cyber work and most research biology, with blocked cyber queries falling back to Opus 4.8 and biology to Opus 5. Anthropic reports about 55% fewer blocks than Fable 5.1. A refusal returns HTTP 200 with stop_reason: "refusal", so an error-rate dashboard won't show it. Refusals before any output are billed in the bio, frontier_llm and reasoning_extraction categories and count against rate limits in every category.

What Security and Life-Sciences Teams Get

Mythos 5.2 stays behind Anthropic's verification programs, with one exception that matters to enterprise security teams. CNBC reports access runs through the Life Sciences and Cyber Verification Programs, while Claude Enterprise customers can use Claude Security, which runs on Mythos 5.2, without joining.

The Cyber Verification Program requires data retention, aims to send a decision or a request for more information within seven business days, and typically adds about five business days before Mythos appears on cloud platforms. On Bedrock, the program is limited to organizations holding a retention exemption for Fable 5.2 or 5.1. Mozilla says Mythos 5.2 "roughly doubled the vulnerability discovery metrics" of Mythos 5.1 in its scanning campaigns, and Implicator points out that the result covers Mozilla's own scanning.

---

What to Do Before You Change a Default

This Week:

1. List every workspace running under ZDR and have your privacy lead decide, in writing, which data classes may enter a 30-day log. Do this before anyone adds claude-fable-5-2 to a LiteLLM or OpenRouter config.
2. Set effort explicitly on every Claude call in your gateway, so a model swap cannot change it for you.
3. Pull 30 recent tasks where Claude Opus 5.5 failed at high or xhigh effort. That set is your Fable 5.2 candidate list. If you cannot fill it, you have no case for moving yet.

This Month:

1. Run those tasks through an effort sweep: Opus 5.5 at medium, high and xhigh against Fable 5.2 at low, medium and high. Record cost per accepted result, and count early stops as failures. Anthropic's guidance is to start Fable at high and step down only where your evals show quality holds.
2. Build escalation as one-way routing, Opus first and Fable on failure, and alert on stop_reason: "refusal" separately from errors.
3. Ask your Anthropic account team three questions about EFS: your place in the beta, whether your use counts as an internal business application, and who staffs the flag review on your side.

Before Moving Production Traffic:

1. Compare your peak token volume with Fable 5.2's rate limit at your tier on Bedrock or the Claude API, and run Anthropic's prefix check test if your code builds the messages array itself.
2. If you run Opus 5.5 at Max, re-test the half-cost claim on your own tasks first, since that comparison matches your setup.

The Bottom Line

Anthropic is working through the two objections that held Fable back. In September, The Register reported that retention terms were a competitive disadvantage and that the market was reluctant to pay a premium for Fable. Fable 5.2 cuts the price by a fifth and puts customer-held monitoring in private beta, weeks after EFS was announced, as the company prepares for a possible IPO as soon as next week.

For routine work, Anthropic's own recommendation is still Opus 5.5 at half the list price. Run the sweep on the 30 tasks Opus failed, then find out which of your workspaces would return a 400 on claude-fable-5-2.

Continue Reading

• Fable 5 Needs Retention On. ZDR Was Never Zero.
• Use Bedrock's None Mode to Enforce Zero Data Retention
• Claude Sonnet 5.5 Costs More Than Opus 5.5 at the Same Score
• Claude Opus 5.5 Rejects the Forced Tool Calls Opus 5 Accepted
• Anthropic Cut Cache Reads 75%. Opus 5 Still Undercuts It.
• Anthropic Adds Red Team Access to Cyber Verification Program