Anthropic this week unveiled Claude Fable 5, its most capable generally available model in the Claude family (a tier above the Opus-class models, in the ‘Mythos-class’). Fable 5 is already attracting significant attention across the legal sector because of its high performance on knowledge work and complex tasks that sit at the heart of legal work.
The model has been made available through a number of AI platforms, including legal AI unicorn Harvey, which announced immediate early access for customers. Harvey says that Fable 5 achieved the highest score yet on its Legal Agent Benchmark (LAB), outperforming previous frontier models on complex legal workflows. According to Harvey, the model also achieved a new high score on its BigLaw Bench evaluation framework.
For the legal sector, the launch represents another step change in the capabilities of the underlying foundation models that power many of today’s legal AI applications. Every major leap in foundation model performance effectively raises the ceiling for what those applications can achieve. Fable 5 appears to be one of those moments.
Early testing suggests that the model performs particularly well on large-scale, multi-document reasoning. Anthropic says Fable 5 is designed to handle large volumes of information and longer running tasks than previous Claude models. That is directly relevant to legal use cases such as due diligence, investigations, contract analysis and regulatory reviews, where lawyers must synthesise information across multiple documents and sources. On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 has the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving.
There is, however, an important caveat or red flag. In fact there’s more than one. First, Anthropic has warned that a model of this capability presents cybersecurity risks and could be misused. On some topics, you’ll receive a response from Opus 4.8 as a form of guardrail.
Secondly, Harvey has highlighted that Fable 5 comes with different data privacy and security arrangements from those typically associated with enterprise legal AI deployments. According to Harvey, Anthropic may retain inputs and outputs for ‘safety reasons’, with data generally deleted after 30 days, and processing taking place in the United States. As a result, access is currently being offered on an opt-in basis, allowing customers to assess whether the trade-off between capability and governance aligns with their requirements. This will be a giant no on client work. Microsoft has notably restricted its employees from using Fable 5 over data retention concerns.
If the past couple of years have been about law firms scrambling to adopt generative AI, this year is about them scrambling to fill the governance gaps that are emerging and ensuring that they remain firmly in the driving seat. Make sure you sign up to receive Legal IT Insider’s brand new report, out next week, that takes a deep dive into four layers of governance that need to be at the top of the agenda: https://legaltechnology.com/liti-to-launch-its-2026-genai-part-3-report-governance-sign-up-here/
