News You Can Use

Anthropic released Claude Fable 5 on 9 June, a generally-available, safety-hardened Mythos-class model. On Harvey's Legal Agent Benchmark it scored 13.3%, an all-time high, up from Opus 4.8's 10.4%, with the competing frontier models well behind on the same test; it also took 93.4% on the easier single-step BigLaw Bench. Three days later, on 12 June, a US Commerce Department export-control directive ordered Anthropic to suspend all access by "any foreign national, whether inside or outside the United States," including its own foreign-national staff, citing a reported method of jailbreaking Fable 5 that the government understood could unlock Mythos's cyber capability. Anthropic disabled both models for everyone worldwide. Anthropic publicly disputes the basis, arguing the jailbreak was narrow (one specific instance) rather than a universal defeat of Fable 5's safeguards.

Two weeks after Opus 4.8 became the first model to break 10% all-pass, Fable 5 had already reset the bar, and within seventy-two hours it was gone. The lesson learnt was that the most capable model can be the best legal tool you have on Tuesday and unreachable by Friday, for reasons out of your control. Do not anchor a build, a pricing model or a client commitment to a single model; design for substitutability. There was the usual race to announce from leading vendors about rolling out Fable which they will now have to backtrack on. Despite being truly brilliant, still scoring just 13.3% on complex agentic workflows shows how these models aren't quite ready for full delegation. But performance on standalone tasks is beyond any other frontier model. For an innovation function the takeaway is important and something we have focused on - model-agnostic architecture, a verification layer that does not rely on one approach, and a procurement posture that treats frontier access as a variable to plan around.

Per-token prices are falling, but the tasks lawyers now run (more reasoning, more agentic multi-step work, more context pulled in) are getting far more token-intensive far faster, and the newest top models are getting more expensive, not less, so per-matter cost is climbing and becoming unpredictable. Kirkland confirmed their platform will run on Palantir's AIP ontology layer, structuring the firm's own institutional knowledge across the PE fund-formation lifecycle, exclusive and not for resale, with a separate hint that Kirkland may fine-tune open-source models into a firm-owned model. Law firms struggle to fund long-horizon AI infrastructure (every dollar of capex is a dollar a partner does not take home). The cost debate even christened a new word "tokenmaxxing" - which carries the risk of reimporting the billable hour's original sin (measuring consumption, not value). Just because something is measurable does not necessarily mean it should be used as THE metric - "a token is the unit a model uses to bill you, it is not the unit your work is measured in". Figuring out the infrastructure, caching, correct model, per-matter usage monitoring is ordinary management discipline that must go alongside any internal build.

This is the cost side of the spend-versus-proof argument, and it reframes the buy-vs-build question as a capital-structure question. Most firms cannot quickly invest heavily even if they want to (the partnership cannot easily write that cheque), which is why the firms that do are repricing partner time into the build rather than raising fresh capital. The value is still in driving adoption and 'activation' of AI into workstreams, rather than the build or buy itself. This makes the market case for an AI-native services layer that operates the tools rather than just selling them. The measurement problem is real - tokens are a consumption meter masquerading as a value metric - while the cost-control problem is mostly solved engineering. The genuinely unsolved bit is pricing the value to the client when your input cost is a volatile consumption meter and your output is unmonitored. For an innovation function, token economics is now an important line in AI business cases: measure spend by matter, right-size models to task difficulty, and treat a defensible private-knowledge layer as the thing worth owning, because the model is rented and increasingly volatile but the institutional knowledge is yours.

Ethan Mollick, whose bestseller Co-Intelligence made the case for humans and AI as collaborators, has walked this framing back. He now argues the assistant-in-dialogue model is giving way to autonomous agents that are "sometimes better than you, and sometimes hilariously worse" across the same jagged frontier, and that the new human skill is judgment over delegation rather than prompting: "when should you refuse AI's help, even when it is offering? When should you hand over the keys entirely?" Microsoft Research's New Future of Work Report 2025 finds enterprise users self-report saving 40-60 minutes a day, but 40% received "workslop" (polished-looking, useless AI output) in the past month, roles are shifting from doing the work to guiding, critiquing and improving it. Without training, employees risk losing core cognitive skills, with entry-level roles the easiest to automate and therefore the first to erode. A Microsoft Research study of around 5.5 million anonymised M365 Copilot sessions shows usage maturing from chat-as-search toward content creation, unevenly across occupations. The scarce resource in legal work was never information but human cognition, and what AI collapsed is the cost of producing words, not understanding them, so "understanding legal work remains as expensive as ever... the legal industry charged exclusively based on producing the work." Roughly 80% of Anthropic's own code is now written by its models, how long until this shift starts coming into legal work as well?

The doing-to-guiding shift maps almost one-for-one onto the legal training pipeline. This is the area in which law firms need to adapt quickly and will make supervisors uncomfortable. If entry-level work is what AI eats first, and entry-level work is how juniors build the tacit judgment the profession runs on, then deskilling is a direct threat to the talent supply that makes a firm worth hiring, and worth treating as a strategic risk in its own right. David Wang's line that we "only got away with billing for production, not understanding" is the verification-tax argument restated from the knowledge-work side. Production is cheap now, understanding is not, and the pricing model has not caught up. The difficulty will be convincing clients to pay for judgement and knowledge rather than input - something that the industry has not yet figured out how to price. Benedict Evans argues that trying to predict which jobs AI displaces is mostly futile because we cannot know how the jobs themselves will change, which is the right corrective to the alarmist headline jobs numbers. The action for an innovation, L&D and knowledge function is concrete: make verification workflows and structured training a deliberate investment in the firm's long-term capability, and decide intentionally when a lawyer should refuse the AI's help and when they should hand over the keys, rather than letting that line get drawn by default (or through laziness).