News You Can Use

OpenAI launched GPT-5.4 on 5 March - the first general-purpose model with native computer-use capabilities. It can take screenshots, navigate applications, use a mouse and keyboard, and execute multi-step tasks with minimal human intervention. 1M token context in the API. 33% fewer false claims versus GPT-5.2. It scores 75.0% on OSWorld-Verified, surpassing human performance at 72.4%. This arrived alongside Google's Gemini 3.1 Pro (dominates 13 of 16 major benchmarks, 1M context), Claude Sonnet 4.6, xAI's Grok 4.20 (four-agent parallel collaboration), and MiniMax M2.5 from China rivalling Claude Opus 4.6 at lower cost. Microsoft launched Copilot Cowork - a new agentic layer in M365 that builds plans and executes them across Outlook, Teams, Word, and Excel, integrating both its own models and Anthropic's Claude. A new E7 licensing tier at $99/user/month bundles everything from May. Andrej Karpathy declared the coding era over. Vibe coding is already "passe," replaced by "agentic engineering" where AI agents build entire applications autonomously.

The significance is not any single model - it is the convergence. Every major lab now offers million-token context, native tool use, and agentic capabilities. The moat has shifted from model quality (largely commoditised) to ecosystem, integrations, and trust. For legal, the practical question is: what does native computer use mean for document-heavy workflows? A model that can see your screen, navigate between applications, and execute multi-step tasks is a different category entirely from a chatbot that drafts text in a window. The Microsoft Copilot Cowork launch makes this concrete for anyone on M365 - describe an outcome, and the system builds a plan and runs it across your Office apps. The new E7 tier signals that Microsoft sees agentic AI as a standard enterprise capability, not a premium add-on. For firms evaluating AI strategy, the message is clear: stop comparing individual models. Start thinking about which ecosystem your workflows plug into.

Lawhive, backed by Google Ventures with approximately £40m in funding, acquired Woodstock Legal Services, a property law consultancy. This is the first time an AI legal platform has bought a traditional UK law firm. Lawhive's AI assistant "Lawrence" handles paralegal-level tasks. In the same fortnight, LawFairy - only the second "technology-only" law firm authorised by the SRA - went live in immigration law. Its founder, ex-Hogan Lovells partner Raj Panasar, made a pointed distinction: probabilistic AI is "fundamentally unsuitable for regulated legal work." LawFairy uses a deterministic, rules-based engine with zero hallucination risk by design. And an AI-powered employment law firm successfully negotiated a £30k settlement for a client - a concrete outcome, not a pilot.

Three different blueprints for disruption, each asking a different question of the traditional legal model. Lawhive's approach is acquisition-led: buy the practice, insert the technology, scale. LawFairy's is architecture-led: build a system that cannot hallucinate, then apply it to complex regulated work. The employment firm is outcome-led: AI does the work, the client gets the result, and nobody asks how many hours it took. For traditional firms, the competitive question has shifted from "will AI disrupt us?" to "which model of disruption gets traction first?" The SRA's willingness to authorise technology-led firms suggests the regulatory environment is evolving faster than many expected. Sequoia's recent "Services: The New Software" thesis maps directly onto this: the next wave of value accrues to platforms that sell completed legal work, not tools for lawyers to work slightly faster.

Jamie Tso's Substack piece "The Origin Story of a Legal Quant" makes the sharpest case yet for a new archetype. These are lawyers who have moved beyond prompting into genuine systems building - agent workflows, automation pipelines, purpose-built tools. His argument is that firms structurally cannot value these people. The billable hour rewards throughput, not leverage. A lawyer who automates a three-week process into an afternoon creates enormous value and gets rewarded with less revenue. Tso proposes a "Jane Street for law" model: a curated network of technically elite lawyers connected directly to clients, with purpose-built infrastructure rather than generic SaaS. He is building LegalQuants to do exactly this from Hong Kong. Separately, Zack Shapiro's "The 10x Lawyer" describes what this looks like in practice: AI amplifies excellent judgment into exceptional output and poor judgment into faster mistakes. The barbell effect is real - elite AI-native practitioners commanding premium rates at one end, commodity workers competing on price at the other. The vulnerable middle is traditional associates and legal tech startups that solve problems these lawyers can now solve themselves. Jack Shepherd's honest account of going from vibe-coding sceptic to believer in two days maps the journey: six stages from "AI as proofreader" to "AI as development team." The LawDroid Manifesto profiled a "Claude-native lawyer" whose prompts average 2,000 words - briefing AI like a senior partner briefs a trusted associate.

The anti-SaaS argument here is worth sitting with. Tso's claim - that "I could build this better" describes capability, not a business model - cuts against the instinct most technically-minded lawyers have. But the market mispricing is real: if the going rate for "AI Architect" roles is $150k, the market genuinely does not know how to value this combination of judgment and technical skill. For firms, there are two questions. First, can you retain these people before someone else gives them a better deal? And second, can you create the conditions for more of your team to become this? The Shepherd piece suggests the tools are now accessible enough that the barrier is not technical ability - it is mindset. The lawyers who are pulling ahead are the ones who stopped waiting for the innovation team to build something and started building it themselves. That is not a criticism of innovation teams. It is recognition that the best tools are often built by the people who understand the problem most deeply. For anyone reading this who has been curious about what Claude Code or Cursor can do - this is the fortnight to find out.