Claude Fable 5.1: What Anthropic's New Flagship Changes for Lawyers

The Release, in One Paragraph

On September 1, Anthropic released Claude Fable 5.1, the successor to the Fable 5 flagship it launched on June 9 (Anthropic suspended access to Fable 5 for all users from June 12 to July 1 after the US government applied export controls to it), alongside Mythos 5.1 (the same model with more permissive safeguards for vetted organizations, available only through Anthropic's trusted-access programs). The version number undersells it. On Anthropic's published benchmarks the new model roughly doubles Fable 5's scores on some agentic tests (52.6% vs. 24.7% on Terminal-Bench-Science 0.1, and 31.4% vs. 17.1% on AutomationBench), and early-access partners reported it working unattended for hours, including one 38-hour machine-learning run. But if you run a law practice, the benchmark story is the third most important thing about this release. The first two are about confidentiality and friction.

Change One: The Data-Retention Asterisk Is Loosening

When Fable 5 launched, it came with a policy no earlier Claude model had: prompts and outputs were retained by Anthropic for 30 days for safety work on every platform where the model was offered, and organizations with zero-data-retention arrangements had to turn retention on to use it unless Anthropic expressly authorized otherwise. Anthropic commits that retained data is not used for training without the customer's express permission, human review of retained content is limited to material flagged by its safety systems and is logged, and deletion after 30 days is automatic except for flagged content or data it is legally required to keep, but it was still a retention stream outside your control, and we advised treating it as a real consideration for matters with strict confidentiality requirements: protective orders, unusual sensitivity, client-imposed data terms.

With Fable 5.1, Anthropic announced Enterprise Frontier Safeguards (EFS), which store safety-monitoring data in cloud infrastructure the customer controls and give privacy equivalent to zero data retention. EFS is rolling out to customers in phases starting this fall. Until then, eligible customers can use Fable 5.1 (and Fable 5) with zero data retention. The retention policy itself has not gone away: Anthropic's help center still lists Fable 5.1 as a covered model, a category whose prompts and outputs are retained for 30 days, and its API documentation says zero data retention is not available for it unless Anthropic expressly authorizes it. The practical translation for a firm: the Fable retention caveat now depends on your contract, not just the model. If your organization is eligible for zero data retention or EFS, the flagship can sit under the same data terms as the rest of the lineup. If not, the 30-day retention still applies. Before retiring a rule that routes strict-confidentiality matters away from Fable, get written confirmation of your organization's eligibility.

Change Two: Far Fewer False Refusals

Fable 5 launched with safety classifiers that can decline requests in high-risk areas such as cybersecurity and biology, and Anthropic acknowledges that benign work in those areas can trigger them too. Anthropic says that with Fable 5.1's updated cybersecurity safeguards, Claude Code users can expect around 60 percent fewer interventions per session than under Fable 5's, and that its updated biology safeguards fire 85 percent less often on benign elementary biology and medical questions. Dual-use cyber tasks such as penetration testing and exploit generation are still redirected to Opus models. For legal work this should mean fewer interruptions in matters that touch these subjects professionally, such as cybersecurity counseling and medical questions, though Anthropic has not published figures for other topics. The professional rules have not changed, and neither has your verification duty. The tool just objects to lawful work less often.

Change Three: The Economics Moved, Quietly but Meaningfully

Fable 5.1's sticker price is unchanged ($10 per million input tokens and $50 per million output tokens), but the price of re-reading cached context dropped 75 percent, to $0.25 per million tokens. Anthropic estimates this cuts effective costs by about 25 percent for typical work and up to about 45 percent for highly agentic work. One API billing change cuts the other way: since September 24, a refusal that arrives before any output is billed at normal rates when it falls in the biology, competing-AI-development, or reasoning-extraction categories; refusals before any output in other categories, including cybersecurity, are still not billed. Anthropic also reports that at Low or Medium effort the new model achieves results similar to or better than Fable 5's at much lower cost (Fable 5.1 defaults to Medium effort on claude.ai; on the API the default is High). On a claude.ai subscription you never see per-token prices, and Anthropic has not said how the cache-read price cut, which applies wherever usage is billed by token, affects subscription usage limits. Our long-standing advice against camping in extended flagship conversations gets softer accordingly, though it does not disappear; the flagship remains the most expensive way to do anything routine.

What Does Not Change: How a Firm Should Route Work

Every model release tempts users to promote the new flagship to daily driver. Resist it, for the same reasons as before. The pattern that serves law firms has three tiers, and 5.1 makes the top tier better without changing the structure:

• Daily volume (drafting, correspondence, summaries, reformatting) belongs on Sonnet, which is fast, capable, and gentle on your usage.
• Serious analysis (contract review, multi-factor tests, long-document reasoning) belongs on Opus, which remains the default for work where the thinking is the deliverable. Anthropic's own Fable 5.1 documentation takes the same line: start most workloads on Opus, and use Fable 5.1 for demanding reasoning and long-horizon agentic work.
• The hardest ten percent belongs to the flagship, used deliberately: take the genuinely difficult problem to Fable 5.1 for a turn or two of strategy (the approach, the tradeoffs, the plan), then carry that plan to a cheaper model for execution. Anthropic says Fable 5.1 is stronger than Fable 5 at long-running agentic work and multistep research, and recommends it for demanding reasoning, which makes the pattern more valuable, not less.

Our full task-by-task breakdown lives in which Claude model to use for legal work, updated for the new lineup.

What About Harvey and the Legal-Tech Stack?

Dedicated legal platforms set their own timelines for adopting new models, and the largest of them are not built on Claude alone: Harvey, for example, offers models from Anthropic and Google alongside OpenAI's. If your firm evaluates Claude against Harvey, the calculus is unchanged: the platforms differentiate on workflow and integrations, not on exclusive access to a smarter brain: Harvey itself offers frontier models from Anthropic, Google, and OpenAI. What a release like this does shift is the floor: capabilities that justified enterprise platform pricing a year ago increasingly ship in the tools a solo practitioner already pays for.

What to Actually Do This Week

• Check which model your tools call. On claude.ai, choose Fable 5.1 in the model picker when you want the flagship. If your firm has API integrations or pinned workflows that call Fable 5, have whoever maintains them review Anthropic's documentation before switching: moving from Fable 5 to Fable 5.1 includes breaking changes, such as forced tool use now returning an error.
• Revisit the Fable retention caveat if your firm policy adopted one, but retire it only after confirming that your organization is eligible for zero data retention or Enterprise Frontier Safeguards. Without that eligibility, Fable 5.1 is still subject to 30-day retention.
• Re-test one hard workflow. If a matter defeated the flagship in July, it is worth one more attempt; the long-horizon gains are the headline capability change.
• Do not rewrite your prompts. The prompting guidance that arrived with this model generation still holds, and if your templates predate it, the bigger win is pruning them: see why over-prompting hurts current models.

The Bottom Line

Fable 5.1 is a rare release where the fine print matters more than the benchmarks for our audience. The capability gains are real but mostly land in long-autonomy work few firms run today. The retention change and the safeguard tuning, by contrast, ease the two practical frictions that made the flagship an awkward fit for a law practice, though for now the retention relief depends on whether your organization is eligible for zero data retention. Use it the way you would use your sharpest partner: not for everything, but for the problems where nothing else will do.