March 2026 did not produce one neat launch day. It followed a rolling wave of releases and previews: OpenAI announced GPT-5.4 on March 5, following Anthropic’s Claude Sonnet 4.6 on February 17 and Google’s Gemini 3.1 Pro on February 19. MiniMax M2.5 and Zhipu’s GLM-5 also emphasized practical agent work. This article considers that early-March snapshot; rumors about DeepSeek V4 and community reports about Grok 4.20 should be distinguished from documented vendor releases.

The common theme in those announcements is delegated work: coding, tool use, document analysis, and longer tasks. That matters because the market is moving beyond chatbot novelty toward systems that can take on more of a workflow. Whether they do that reliably is a question for evaluation, not launch-day enthusiasm.

The technical story: from chat to agentic work

The most important technical development is not that these models got incrementally smarter. It is that they became more operational. GPT-5.4 combines improved factuality with native computer use, tool search, and up to 1 million tokens of context, making it less like a conventional chatbot and more like an agent that can plan, act, and verify across long workflows. Claude Sonnet 4.6 pushes in the same direction with stronger computer use, long-context reasoning, and agent planning, also with a 1 million-token context window in beta. Gemini 3.1 Pro is expressly framed by Google as upgraded “core intelligence” for complex problem-solving across consumer, developer, and enterprise products. GLM-5 emphasizes long-running agent tasks, while MiniMax says M2.5 was trained in real-world environments for coding, search, office work, and tool use.

The breadth of these releases also raises a structural market question: how easily can an organization switch models without replacing its workflows? For lawyers, competition involves more than a benchmark ranking. Vendor concentration, portability, cloud distribution, and platform defaults can all affect bargaining power and the practical availability of alternatives.

The practical story: where lawyers will actually feel it

For law firms, legal departments, and courts, the practical question is which workflows can benefit from these systems and how to verify the result. OpenAI reports a 33% relative reduction in false individual claims compared with GPT-5.2 on a particular set of prompts flagged for factual errors. Harvey reports 91.0% on its BigLaw Bench suite. These are evaluations under specified conditions, not estimates of accuracy on every legal assignment. Databricks, quoted in Anthropic’s announcement, reports Sonnet 4.6 matching Opus 4.6 on OfficeQA, while Google’s announcement describes Gemini 3.1 Pro across its API, Vertex AI, Gemini app, and NotebookLM.

Where the system runs also matters for legal work. Local software, hosted models, connected apps, and mobile devices create different data flows and operational constraints. A firm should identify where client information goes and who can access it before treating a new interface as a secure workflow.

One immediate legal issue is authorship. The Supreme Court docket records a March 2, 2026 denial of review in Thaler v. Perlmutter; that denial was not a new Supreme Court ruling on the merits. The Copyright Office’s 2025 report distinguishes human-authored expression from material generated by AI and explains that using AI as an assistive tool does not itself bar copyright. For lawyers, the practical questions include what the human contributed, what material is licensed, and which rights a proposed use may affect.

Next is privacy and governance. The EDPB’s opinion on AI models addresses personal data, legal bases, and the consequences of unlawful processing; California’s privacy agency likewise describes purpose-limitation and data-minimization duties for covered businesses. These questions belong in system design. The EU AI Act follows a phased timetable with different obligations and exceptions, so readers should consult the Commission’s implementation timeline for the relevant system and date. NIST’s February 2026 AI Agent Standards Initiative also makes clear that agent identity, security, and interoperability are active areas of standards work. For legal teams, the practical task is to connect those requirements to data flows, human review, testing, recordkeeping, and vendor controls.

Then there is liability. Responsibility will depend on the governing law and the facts, including what was promised, how a system was used, and what review was reasonable. The FTC’s Workado proceeding illustrates scrutiny of unsupported AI accuracy claims. Rytr requires a different account: the FTC reopened and set aside its 2024 consent order on December 22, 2025, before this article’s original publication. Contract terms remain central to allocating confidentiality obligations, training-use restrictions, audit rights, service levels, and indemnities. In many organizations, an AI dispute may begin with those commitments before it reaches broader regulatory or tort questions.

Key takeaways

  • March 2026’s launch cycle shows that the frontier has shifted from chat quality to agentic execution, long-context reasoning, and document-heavy work.
  • Chinese challengers such as GLM-5 and MiniMax M2.5 are no longer peripheral. They are part of the pricing and governance story now.
  • The biggest legal pressure points are IP, privacy, regulatory oversight, fairness, liability, and competition, not just benchmarks.
  • Lawyers who understand model launches only as tech news will miss the larger point: these releases are already rewriting the economics and risk architecture of practice.

Lawyers should treat frontier-model launches the way they treat major appellate decisions: not as curiosities, but as signals. The firms and legal departments that follow these developments closely will help shape the next generation of AI governance standards, licensing terms, regulatory arguments, and liability theories. The ones that do not will inherit rules written by others. Staying informed on AI is no longer optional for lawyers. It is rapidly becoming part of competent practice.