← The Intake
The Intake · Weekly briefing

This week in AI, for legal

Anthropic's Claude Opus 5 arrived in Harvey this week with a genuine jump in legal-accuracy benchmarks, three weeks after OpenAI launched its GPT-5.6 family and then, on July 30, cut two of its three prices by up to 80%. Legora added a fifth acquisition to a buying spree that started in March and picked up a 550-job, $87.5 million research hub in Manhattan along the way. And AI's copyright reckoning landed on two continents in the same week: Delhi's High Court ruled that training a model on copyrighted news counts as fair dealing, while Munich's regional court was due to rule this morning on whether training on copyrighted music requires a licence at all.

Week of 25 – 31 July 2026
Category Market intelligence
Reading time 9 minutes
01 — The week at a glance

Five stories that matter

July 24 — Frontier models
Opus 5 scores 11.7% all-pass on Harvey's own Legal Agent Benchmark, up from the prior Opus model, while generating 26% fewer tokens at a comparable reasoning level. It's rolling out to Harvey's US customers now, with EU and Australian availability to follow.
July 24 — Regulation
Justice Amit Bansal declined ANI's injunction against OpenAI on a prima facie finding that training for research falls within India's fair dealing exception. The underlying case, over whether ChatGPT reproduces ANI's reporting or fabricates quotes attributed to it, continues.
July 27 — Legal AI
Governor Hochul's announcement puts a Stockholm-founded legal AI company at a nearly 100,000-square-foot office at 11 Madison Avenue, staffed for engineering, product, and legal engineering work.
July 29 — Legal AI
Wexler's engine isolates who said what, to whom, and why it matters across document sets that can run past a million pages per case. The deal follows Legora's acquisitions of a Canadian agentic platform, a legal research start-up, a regulatory-intelligence tool, and a commercial real estate agent builder, all since March.
July 31 — Regulation
Munich's regional court is due to rule this morning on whether training AI on copyrighted music requires a licence at all
GEMA, representing more than 95,000 German rightsholders, wants the court to require permission before any AI lab trains on protected songs. Under German law a first-instance ruling takes effect immediately, even while Suno appeals.
02 — The platform being assembled

Legora has made five acquisitions since March. Together they look like a platform being bought, not built.

Wexler, the litigation fact-intelligence start-up Legora agreed to acquire on July 29, is a small company: eighteen people, founded in 2023, previously raised under $7 million across a pre-seed and a seed round. On its own it barely registers. As the fifth Legora acquisition since March, following a Canadian agentic platform, a Stockholm legal research engine, a regulatory-intelligence tool, and a commercial real estate agent builder, it reads differently: a company assembling the full width of a legal AI platform through acquisition, at a pace few competitors can match. The deal lands two days after New York state offered Legora up to $10.5 million in incentives to build a 550-job, $87.5 million research and engineering hub in Manhattan, its clearest signal yet that the US, not Sweden or the UK, is where it intends to fight hardest.

5
Acquisitions since March 2026, each adding a distinct capability
$5.6B
Legora's valuation after its $600M Series D, closed earlier this year
550+
New Manhattan jobs planned, backed by up to $10.5M in New York incentives
AcquisitionWhat it bringsWhere it sits in the buying spree
Walter AICanadian agentic legal AI platform1st — market entry into Canada
QuraAI-native legal research and semantic search, live across 27 jurisdictions in competition law2nd — legal research
GraceviewReal-time regulatory and legal intelligence3rd — regulatory monitoring
CadastralAI agent platform built for commercial real estate4th — vertical specialisation
WexlerLitigation fact-intelligence: verifies who said what across 1M+ documents per case5th — litigation support
The question worth asking

Buying five companies in five months is a fast way to fill out a product map. It's a slower question whether the capability that arrives with each acquisition runs on the client's own playbooks and escalation rules, or whether it brings its own defaults that a firm now has to configure around, the same integration work a firm would face buying five separate point solutions itself, just wrapped in a single invoice.

The Flank read

Assembling the full width of a platform through acquisition is a legitimate strategy, and Legora is executing it faster than almost anyone else in the category. But it answers a different question than the one enterprise legal teams are actually asking. Inexpensive work is being done by expensive resources, and no amount of bolted-together point solutions changes who reviews the output before it reaches a client, or whether the work was ever configured against that specific team's templates and fallback positions in the first place. Tools are a commodity. A platform built from five commodities is still a platform of tools, not an outcome.

03 — The moving floor

OpenAI cut its prices 80% three weeks after launch. Anthropic's newest model reached Harvey the same week. The floor under every legal AI product keeps moving.

OpenAI's GPT-5.6 family, three models it calls Sol, Terra, and Luna, went generally available on July 9, after a two-week limited release the company said followed a request from the US government. On July 30, with little fanfare, OpenAI cut Luna's price by 80% and Terra's by 20%, crediting efficiency gains from using the models themselves to help rewrite parts of its own production code. Anthropic's Claude Opus 5, live since July 24, reached Harvey's platform in the same week: an 11.7% all-pass score on Harvey's own Legal Agent Benchmark, a meaningful step up from the prior Opus model, reached while producing 26% fewer tokens at a comparable reasoning level. And the protocol connecting all of this to the tools a lawyer actually touches changed too. The Model Context Protocol's July 28 specification moves from an always-connected, stateful design to a stateless request-response model, the kind of change that lets an agent call a tool without holding a live connection open, and that Anthropic says over 950 MCP servers already support.

Cheaper
GPT-5.6 Luna and Terra
Luna's output price falls from $6 to $1.20 per million tokens, Terra's from $15 to $12. Sol, the flagship, is untouched. OpenAI says the cuts come from efficiency gains, not a discount.
More capable per token
Claude Opus 5 in Harvey
A real jump on Harvey's Legal Agent Benchmark, achieved with fewer tokens than its predecessor needed for similar performance. Strongest so far in transactional and disclosure-heavy work.
Easier to wire together
MCP's stateless spec
Moving to request-response lets agents call tools without a persistent connection, cheaper to run at scale and easier for any vendor to plug in a new model without a rebuild.

None of this happens for free. Microsoft disclosed on July 29 that it signed more than $130 billion in new data-centre leases in the quarter ended June 30 alone, taking its total future lease commitments to $329.1 billion, up from $196.6 billion the quarter before. Somebody is financing the compute that makes a cheaper, better model possible every few weeks. It is not, generally, the legal AI vendor selling access to it.

The question worth asking

If the model underneath a legal AI product gets faster, cheaper, and easier to connect to other tools every few weeks, and if switching which model powers a given workflow requires no visible change for the lawyer using it, what is a legal AI vendor's actual advantage built on: the model, or everything wrapped around it? The honest answer for most products is the latter. Few say so plainly.

04 — The copyright reckoning

India ruled training is fair dealing. Germany may rule the opposite on music, this morning. Both point to the same shift: the fight is moving from training data to outputs.

India's Delhi High Court and Germany's Munich Regional Court have spent the past year hearing structurally similar arguments, whether training an AI model on someone else's copyrighted material is itself the infringing act, and have reached, or are about to reach, very different starting points. On July 24, Justice Amit Bansal declined ANI Media's bid for an injunction against OpenAI, finding on a prima facie basis that training ChatGPT on ANI's news content for research falls within India's fair dealing exception. The underlying case continues, but on a narrower and arguably more consequential question: whether ChatGPT reproduces ANI's reporting or fabricates quotes wrongly attributed to it. Munich's 42nd Civil Chamber, under Judge Schwager, was due to rule this morning on a different theory entirely. GEMA, representing more than 95,000 German rightsholders and over 2 million worldwide, wants the court to find that training on six specific songs, including Rasputin and Mambo No. 5, required a licence Suno never obtained. A loss for Suno would not just settle six songs. Under German law, a first-instance ruling takes effect immediately, even while an appeal is pending.

Delhi's frame
Training on copyrighted material for research is presumptively lawful. The real legal exposure sits downstream, in whether a model's outputs reproduce protected content or misattribute fabricated statements to the rightsholder.
Munich's frame
Training itself is the alleged infringement. A ruling for GEMA could mean no AI lab may train on protected European music without a licence in hand first, before a single output is ever generated.

Neither ruling touches the EU AI Act's Article 50 transparency obligations flagged in this briefing last week, which still bind in two days regardless of how Munich rules this morning. That is a separate, narrower question about disclosure. What Delhi and Munich are fighting over is bigger: not whether a business has to label AI-generated content, but whether the underlying model was allowed to exist in its current form at all.

The Flank read

Two courts, two different theories of where liability attaches, and neither has finished. For an enterprise legal team procuring AI tools, that is not an abstract jurisdictional curiosity. It is a due-diligence question that now has to be asked jurisdiction by jurisdiction: what was this vendor's model trained on, under whose law, and does the answer change depending on where the output gets used. A supervision layer that reviews every output before it reaches a client doesn't resolve the training-liability question either, but it answers the question a court can't yet: was this specific piece of work checked against this specific client's standard before anyone downstream relied on it. That distinction matters more, not less, while training law stays unsettled.

05 — How much agent is an agent

Harvey's newest features draft, review, and report without a lawyer initiating each step. Whether that counts as agentic, in the sense that matters, depends on who still has to check the work.

Harvey's July product update, published as "The Brief," ships new agentic features it says handle more of the drafting, review, and reporting work, so lawyers can focus on analysis and judgement rather than assembly. Prompt dictation arrived in Harvey's Word and Outlook add-ins the same month, letting a lawyer speak a request rather than type it, and Claude Opus 5 is rolling in alongside. Harvey also deepened its Microsoft relationship: Microsoft's own Corporate, External and Legal Affairs group is now using Harvey's platform internally, while Harvey leans further into Microsoft 365 Copilot for its own operations. Read generously, this is a real step toward autonomous execution. Read carefully, it's also consistent with a faster, better assistant, one that still waits for a lawyer to open it, prompt it, and send what it produces.

The distinction is not academic. An assistant that drafts, reviews, and reports faster still requires a human to initiate every one of those steps and to catch what slips through. An agent, in the sense this briefing has used the word for legal work, picks up a request and completes it end to end, with a human reviewing exceptions rather than performing the work itself. Harvey's own language, "so lawyers can focus on analysis and judgement," is compatible with either reading. Nothing in this month's release notes says which one it is.

1,598
Verified AI-hallucination cases now tracked in US courts, roughly 140 added in recent weeks
~8/day
Pace at which new cases are being documented, up from five to six a day reported in April
The question worth asking

When a vendor ships a feature that drafts, reviews, and reports without a person initiating each step, does it also publish how that output gets checked before anyone downstream sees it, or is the autonomy itself the entire announcement? The hallucination tracker's accelerating pace suggests the industry's verification story has not kept up with its autonomy story, whoever is shipping it.

06 — So what

What this week tells us

A frontier model cutting its own price by 80% weeks after launch, a legal AI company assembling a platform through five acquisitions rather than building one, two courts reaching opposite starting points on whether AI training itself is lawful, and a vendor shipping more autonomous-sounding features without saying who checks them: none of this week's stories were coordinated, and all of them describe the same market working through the same unresolved question. Intelligence keeps getting cheaper and more capable on a schedule now measured in weeks, not years. Legal AI vendors are racing to own the full stack, by acquisition where building would be slower. Liability for what AI produces is being litigated jurisdiction by jurisdiction, with no sign of convergence. And the gap between shipping autonomy and proving it was supervised hasn't closed.

The model layer is now a moving target
GPT-5.6's price cuts, Opus 5's benchmark gains, and MCP's protocol shift all landed in the same fortnight. Whatever model powers a legal AI product today will likely be replaced, repriced, or reconnected within weeks.
Platforms are being bought, not just built
Legora's five acquisitions since March assemble research, regulatory monitoring, a real estate vertical, and litigation fact-checking under one roof faster than any of it could be built in-house.
Copyright liability is fragmenting by jurisdiction
Delhi treats training as presumptively lawful and pushes the fight to outputs. Munich may treat training itself as the infringement. Enterprise buyers now face a genuinely different legal exposure depending on where a model was trained and where it's used.
Autonomy is outpacing proof of supervision
Harvey's new agentic features arrive the same month the hallucination tracker adds cases at its fastest documented pace. Neither vendors nor courts have settled what "checked before it left the building" needs to look like at agent speed.
The Flank view

Every story in this week's briefing returns to the same imbalance: inexpensive work is being done by expensive resources, and nothing here closes that gap on its own. A cheaper, better model doesn't remove the review step between a draft and a client. A platform assembled from five acquisitions doesn't automatically inherit any one client's own playbooks and fallback positions just because it now owns five companies' worth of point solutions. And a court ruling on whether training was lawful, in Delhi or in Munich, says nothing about whether the specific output a lawyer relied on this week was checked before anyone downstream saw it.

Tools are a commodity, outcomes are not, whichever model happens to be cheapest this month, whichever company Legora has bought this quarter, whichever jurisdiction's copyright theory eventually wins out. The spend this actually competes for is the services budget, outside counsel, ALSPs, offshore delivery, internal paralegal headcount, that in enterprise legal typically runs ten to fifty times larger than the software line, because that is where the routine, repeatable work still sits. Outsourcing that work to supervised agents, agents built on your own templates and escalation rules, with a person reviewing every output before it leaves the system, is what turns a week like this one, cheaper intelligence, a bigger platform, an unresolved copyright fight, into something other than more noise about tools. It's the gap between a faster instrument and the work actually getting done.

Subscribe

The Intake

Weekly briefings on what's actually changing in legal AI — the market shifts, regulatory moves, and structural questions that matter for enterprise legal teams. Written by the Flank team.

Subscribe on Substack
Flank

Outsource legal work to supervised agents

Enterprise legal teams use Flank to handle high-volume contracting end-to-end: NDAs, MSA redlines, procurement, triage. Agents that know your templates, terms, and escalation rules. Every output reviewed before it leaves the system.

Learn more at flank.ai