News You Can Use

Edition 44 · 15th - 30th June 2026

News You Can Use

Opening

How much of your legal AI capability do you actually control, and how much are you renting from a handful of foundation model providers who can change the terms without asking you?

The US government switched Anthropic's Fable 5 off worldwide, but it has now returned. OpenAI's newest models, the GPT-5.6 family, launched the same fortnight available to about twenty approved partners and nobody else. And Legora moved its heaviest agent tier onto consumption pricing, passing the raw token bill through to the matter. Access, and the price of access, both moved further out of the customer's hands.

That has restarted a conversation about open weights, self-hosting and fine-tuning. If the frontier model is a sovereign-policy variable and the pricing is volatile, the models you can actually rely on / own start to look more interesting. GLM 5.2, the new open-weight model from China's Zhipu AI, shipped under a fully permissive MIT licence in mid-June and benchmarked just behind Opus 4.8 on coding. A question to ask - is there a case for owning the GPUs and running a capable model inside your own walls? For most firms the honest answer is still no, but it is no longer a daft question.

This edition covers sovereignty, at two levels. At the country level, the US gating and the UK's own moves this fortnight (the government naming legal services as the first proving ground for its AI sandbox, the SRA rewriting supervision guidance to cover delegating work to AI, the senior judiciary warning about over-reliance) all come back to who sets the terms on which legal work gets done. At the firm level the question is narrower and more useful: which parts of your stack do you own, which do you rent, and have you tested what happens when the rented part vanishes or reprices?

Deep Dives

Three stories worth your time

The Off Switch

Anthropic - Redeploying Claude Fable 5|NBC - US lifts ban on Anthropic's powerful Fable 5 AI model|Mishcon de Reya - Who pulls the plug: the Anthropic Fable affair|The Geek in Review - Beyond the Model: The Ownership Problem

What
Anthropic switched Fable 5 and Mythos 5 off worldwide on 12 June under a US export-control directive, after Amazon researchers reported a jailbreak. On 26 June a letter from Commerce Secretary Lutnick lifted the block on Mythos 5 for an "Annex A" list of roughly a hundred named US institutions, government partners, and the foreign-national staff of those bodies and of Anthropic itself. On 30 June the Commerce Department lifted the export controls entirely, and Fable 5 came back globally on 1 July. It returned with an improved safety classifier that reroutes any request matching the reported jailbreak to Opus 4.8, and with a phased pricing return. Mythos 5, the more capable sibling, stays restricted to the Glasswing programme and select researchers pending a broader trusted-access scheme. This was not confined to Anthropic: on 26 June OpenAI previewed its next-generation GPT-5.6 family (Sol for frontier reasoning and long-horizon agentic work, Terra a cheaper everyday model, Luna the fastest) but, at the Trump administration's request, released it only to about twenty government-approved partners, the first time the US government has pre-emptively gated a US model launch. OpenAI made its discomfort public, saying it "believes in broad access" and that it does not think "this kind of government access process should become the long-term default", because "it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them". During the Fable restrictive period, Legion LegalTech, sued the US government, calling the shutdown "immediate, irreparable, and existential".
So what
Fable 5 is back but the outage has people wary. For nineteen days the best model was unreachable worldwide for reasons no customer could influence. With OpenAI's newest models shipping into the same gate the same fortnight, a firm should treat this as the emerging shape of frontier-model access, not an Anthropic anomaly. We have often said to design for substitutability; this fortnight shows why that is a procurement requirement rather than a nice-to-have. Every workflow that matters needs a named fallback model and a tested path to it, availability and sovereignty belong on the risk matrix next to confidentiality, and anything client-facing should run on a tier you can defend keeping (Opus 4.8 with ZDR, not whatever topped the benchmark on Tuesday). Law firms must own the matter knowledge and judgement, but rent the rest with your eyes open. The thing you control is your own data and process, and that is the thing worth investing in, because the frontier model is now something you plan around rather than an asset you hold. It will be crucial to have a back up plan or ensure that we do not become overly reliant on one model, provider or platform.
SourcesAnthropic - Redeploying Claude Fable 5|Anthropic - Statement on the suspension directive|NBC - US lifts ban on Fable 5|The Hill - Trump administration lifts limits on Anthropic's new AI models|TechCrunch - Trump admin releases Mythos to 100+ US companies, agencies|The Next Web - Legion LegalTech sues US over Fable 5 and Mythos 5 shutdown|Mishcon de Reya - Who pulls the plug|OpenAI - Previewing GPT-5.6 Sol|CNBC - OpenAI limits new AI models to trusted partners at US government request|The Geek in Review - Beyond the Model: The Ownership Problem|Stephen Smith - The Best Legal AI Most Lawyers Will Never Get to Use Was Here for Three Days

The Meter Is Running

Legora - Consumption-based pricing|The Lawyer - Firms face hike in AI cost base as Legora gears up for pricing switch|Artificial Lawyer - Harvey trains open-source models to encode law firm workflows|Legal IT Insider - HSF Kramer develops proprietary AI platform with Acora and Microsoft|Thomson Reuters - Future of Professionals 2026

What
Legora moved its heavy agentic tier (Agent Pro) to consumption-based pricing on 23 June, billing for the work the agent delivers, attributable to the matter, with dashboards and spend thresholds; the lighter Legora Agent stays included, and the framing is "from hours billed and seats licensed, to outcomes delivered". Harvey confirmed proof-of-concept work fine-tuning open-source models on individual firms' processes, and Thomson Reuters, opening early access to a rebuilt CoCounsel on 22 June, confirmed it is building its own model ("Thomson", via the Safe Sign acquisition). Two UK firms went public with proprietary builds in a week: Shoosmiths' "Project Apollo" contract-review platform on Azure (24 June), and HSF Kramer's "Sovereign System of Intelligence" with Acora and Microsoft, whose CTO calls the firm's own data "a moat". Against all this, Thomson Reuters' Future of Professionals survey put $143bn of US client revenue "at risk" from the implementation gap, with 78% of corporate clients calling AI-enabled quality gains essential and just 6% saying providers deliver them.
So what
The cost base and the value model are being rebuilt at the same time, and they pull in opposite directions. On cost, the move off per-seat is now industry-wide, and consumption pricing passes the volatile token bill straight through to the matter, which is honest but turns every high-token workflow (DD at scale, contract families, ediscovery, long document agentic workflows etc) into a number a partner has to manage and recover. On value, the TR gap shows that clients want demonstrably better work and almost none think they are getting it, so the firm that can show it (not just do it faster) captures the premium. The strategic response from the firms with capital is the same in every case: build a proprietary layer over your own data and process, because that is the part you own when the model is rented and the price is moving. As the model becomes the commodity, the defensible asset is your matter knowledge and your ability to prove the outcome, not the tool you bought to produce it. Vendors correlating tokens used to value provided are missing the point of legal work, and we have to be careful to build genuine tracking and ROI monitoring for use cases that are token intensive to make sure there is business value.
SourcesLegora - Consumption-based pricing|Law360 Pulse - Legora to move to consumption-based billing|The Lawyer - Firms face hike in AI cost base|Artificial Lawyer - Harvey trains open-source models|LawSites - Thomson Reuters CEO on the next generation of CoCounsel|Legal IT Insider - HSF Kramer proprietary platform|Lawyer Monthly - Shoosmiths Project Apollo|Thomson Reuters - Future of Professionals 2026|Gary Marcus - The month Generative AI lost its mojo

The UK Sets Its Terms

GOV.UK / Legal Futures - AI Growth Lab names legal services first focus area|SRA / Legal Futures - SRA rewrites supervision guidance to cover delegation to AI|Judiciary.uk - Dame Victoria Sharp, "Without Fear or Favour"|Legal IT Insider - Gen AI and the Practice of Law 3 (Neil Cameron)

What
Three UK institutions moved on AI this fortnight. The government's "AI Growth Lab", a supervised regulatory sandbox convening the SRA, LSB, ICO and the Council for Licensed Conveyancers, named legal services as its first focus area, with AI analysis of property sales packs as the worked example and applications opening this summer. The SRA, updating its supervision guidance after Mazur, expanded it from nine pages to twenty-four and for the first time addressed delegating work to AI systems alongside paralegals and trainees, with compliant and non-compliant case studies. And Dame Victoria Sharp, President of the King's Bench Division, used a 22 June Inner Temple lecture to warn that the threat to judicial independence "is not a dramatic coup by machine" but "a gradual drift: standardised prompts, standardised summaries, standardised risk scores, and eventually standardised dispositions", and against "hidden dependencies that shape the answer before the judge has fully reasoned". Neil Cameron's 100-page "Gen AI and the Practice of Law 3" framed the same period's question as "what a firm can defend, to whom, and on what evidence", flagging that a defensible methodology for validating AI-assisted review "does not yet exist".
So what
This is the UK setting its own terms while the enforcement drumbeat continues elsewhere, and it lands closer to home than the US sanctions stories. The Growth Lab choosing conveyancing and legal services as where it will let firms test AI under supervision is a signal that regulated experimentation, not prohibition, is the direction, and firms with something to test should be paying attention to the application window. The SRA guidance is the floor; it does not yet say "the rules don't mention AI" lets you off, because the existing supervision duties already bite. Cameron's defensibility point is the one to internalise: the live governance question has moved past "is the output any good" to "can we show, to a regulator or a court, how we validated it", and the honest answer is that the method for doing that on interpretive AI synthesis is still being invented. Sharp's warning is the long shadow: the risk is not a rogue machine but a slow erosion of independent judgement through over-reliance, which is exactly the deskilling argument from the bench. AI Literacy efforts need to focus on verification work, importance of humans in the loop who understand and engagement with the outputs instead of just approving them and crucially, support people in being transparent in their AI use so we can continue to improve the outcomes.
SourcesLegal Futures - AI Growth Lab and legal services|GOV.UK - AI Growth Lab|Legal Futures - SRA supervision guidance after Mazur|Law Society Gazette - SRA supervision and AI|Judiciary.uk - Dame Victoria Sharp, "Without Fear or Favour"|Legal IT Insider - Gen AI and the Practice of Law 3|ICO - One year on: the Data (Use and Access) Act

Worth Reading

Everything else worth a click

- Market Moves

Lawyer Monthly - Shoosmiths Launches Project Apollo

A UK national firm's own firmwide contract-review platform on Azure, pitched on explainability (showing why amendments were made) and surfacing the firm's know-how so juniors learn faster; pilots claimed a ~40% cut in first-pass review time. The build-vs-buy story below the Magic Circle tier.

Law.com - Eudia Partners With Microsoft

Azure deployment plus M365/Teams/Copilot distribution. Distribution through the Microsoft stack is how legal AI reaches in-house teams already living in M365; the more strategic of the day's vendor moves.

Mistral - OCR 4

A self-hostable, structure-aware document extractor at $4 per 1,000 pages, 170 languages, single-container on-prem. Document ingestion is the front door of every DD, RAG and contract pipeline, so a cheap EU-vendor OCR that runs inside your own walls matters for any firm with data-residency constraints.

Cooley - GO Lab, Powered by Legora

Cooley launched a founder-facing AI legal portal on Legora's white-label "Portal" product, debuting with Y Combinator's summer cohort. A top-tier firm productising its know-how into a client-facing channel under its own brand, a stickier surface than seat sales.

Artificial Lawyer - Summize Acquires InnoLaw Group

A UK CLM vendor buys a consultancy's people and IP, the read being that contracting success is now an implementation and change-management problem rather than a model problem. CLM-layer consolidation continues.

tech.eu - Jupus Raises €13m Series A

Led by Semapa Next, a European funding datapoint pitched as "the next generation of AI-driven law firms". The capital is still flowing into the in-house and SME-firm layer.

BARBRI Acquires Lega

A legal-education company buys an AI-competency workshop platform and names Christian Lang Head of Innovation. Consolidation moving into the AI training and upskilling layer, which is where adoption (the actual bottleneck) gets solved.

- Models and Big Tech

Z.ai / Zhipu - GLM 5.2, "Mythos at Home"

A 743B-parameter open-weight model (weights out 16 June under a permissive MIT licence, no regional restrictions, 1M-token context) that benchmarks just behind Opus 4.8 on coding and, per Semgrep, beats Claude on some cyber tasks at a fraction of the cost. The credible open-weight alternative a firm could actually self-host, and the reason "should we own the GPUs" is now a real question (see intro). Provenance is Chinese, which is its own governance call. Semgrep - We have Mythos at Home.

OpenAI - Previewing GPT-5.6 Sol

OpenAI's next-generation family (Sol for frontier reasoning and long-horizon agentic work, Terra a cheaper everyday model, Luna the fastest) previewed on 26 June but, at the Trump administration's request, restricted to ~20 government-approved partners under the 2 June federal-benchmarking executive order, the first pre-emptively gated US model launch. OpenAI objected that such a process should not "become the long-term default" (see DD1). The sign the trusted-list regime now spans both American frontier labs. Axios | CNBC.

Vals AI - Harvey Legal Agent Benchmark Refresh

Under the strict all-pass standard, the best model now completes 11.25% of full agentic legal tasks (Claude Fable 5), with Opus 4.8 at 9.58% and the rest well behind. The honest counterweight to "AI beats lawyers": even the leader finishes barely one full legal task in nine end-to-end.

Crosby - Multi-Turn Negotiation Bench

A benchmark scoring models on negotiation as a sequence of judgement calls. Frontier models cluster at 44-50%; humans still beat them at finding new routes to resolution, while models anchor to their opening position. Useful framing for where agentic negotiation actually is.

Gary Marcus - The Month Generative AI Lost Its Mojo

The commoditisation thesis sharpened: LLMs becoming a commodity, OpenAI reportedly leaning toward delaying its IPO, and the warning that "hyperscaling could prove to be among the biggest financial blunders in history". The bubble deflating slowly rather than popping. The macro backdrop to DD2.

- Adoption and Practice

Anthropic - Agentic Coding and Persistent Returns to Expertise

A study of ~400,000 Claude Code sessions finds a clean division of labour (people decide what to build, the agent decides how) and that domain expertise, not tool proficiency, is what amplifies the tool. The steering skill that survives is a function of domain command, and legal is among the fastest-growing non-software user groups.

Microsoft Research - How Copilot Changed the Pace of Work in Word

A causal study of 72,000+ Word users finds adopters complete comparable work ~40-45% faster after twelve weeks, with a spillover onto non-Copilot tasks. Careful and independent (the authors stress measured time is not productivity), which makes it the measurement-rigour foil to vendor ROI decks.

OpenAI - The Shift to Agentic AI: Evidence from Codex

The first large-N evidence that agentic adoption is a token-share story, not a user-count story. Legal sits at the back of the queue on average but intensive where it lands; OpenAI's own legal function went from near-zero to ~75% token share in about three months once friction was removed. Cite carefully (it is a coding tool, OpenAI's own staff), but the friction-removal arc is the enablement thesis in miniature.

RSGI - The Accelerating Impact of Legal AI

A survey of 136 Harvey customers: 55% now call the tool "foundational", 77% justify licences by better outcomes, and there is hard insourcing evidence (a Spanish M&A run in-house against ~€200k external quotes). Read as "among sophisticated adopters", never a market rate, but the pricing disconnect (firms discuss AI efficiencies, in-house barely sees it) is the live client-value story.

Microsoft - Word Legal Agent Goes Worldwide Public Preview

The Copilot legal agent in Word (contract review, risk flags, playbook comparison, native tracked-change edits) expands to worldwide Public Preview, GA early July. A competent redline agent inside the tool every lawyer already uses, and the direct threat to point-solution redline vendors.

Non-Billable - Five Takeaways From LegalTechTalk London

Tool overload, general-purpose AI squeezing specialists, governance over policy PDFs, career-model disruption, and Ben Allgrove's reusable line: judge AI against real historical practice, not perfection. UK conference colour with one quotable artefact.

- Regulation and Courts

Legal Futures - AI Growth Lab Names Legal Services First

The government's supervised sandbox picks legal and conveyancing as its first proving ground, convening the SRA, LSB, ICO and CLC, with applications opening this summer (see DD3). The state choosing regulated experimentation over prohibition.

Legal Futures - SRA Rewrites Supervision Guidance to Cover AI

Post-Mazur, the guidance grows from nine to twenty-four pages and, for the first time, addresses delegating work to AI systems with compliant and non-compliant case studies. The closest the UK has come to concrete AI-supervision expectations, though framed as a supervision update rather than "the SRA's AI rules".

ICO - One Year On: the Data (Use and Access) Act

A twelve-month stocktake (19 June), flagging the statutory AI/ADM code of practice the ICO must now produce. The DUAA's reform of automated-decision rules plus the forthcoming code is the UK's nearest thing to AI-specific data law.

Sidley - EU Digital Omnibus Defers Key AI Act Deadlines

Parliament adopted the Omnibus on 16 June and the Council gave final sign-off on 29 June, deferring standalone high-risk obligations to December 2027 while transparency/Article 50 marking stays live at August 2026. For UK firms advising EU-exposed clients, the deadline relief is the practical planning point (in force on Official Journal publication, expected shortly).

LawSites - Third Circuit Hears Thomson Reuters v ROSS

The appeal in the long-running training-data fair-use case (26 June). The leading US precedent on whether training legal AI on copyrighted material is fair use, and the outcome shapes the vendor risk posture UK firms inherit.

- Critical Perspectives

Stephen Smith - Your AI Chats Probably Aren't Privileged. Good. They Don't Need to Be

Firms using enterprise AI with proper contractual terms and matter-level controls are protecting client information the right way; the fix is buy enterprise terms and configure them, not fear privilege. A clean adoption message against the consumer-tool scare stories. His companion piece argues new state AI rules "don't create new duties, they just took away your best defence".

Jordan Furlong - The Unbundling of Lawyer Institutions

AI strips firms and law schools of their commodity features; their future depends on rebuilding around the highest-value judgement and assurance work that survives. The structural counterpart to DD2's "own the matter" argument.

Ed Zitron - Cargo Culture

The industry mindlessly copying past infrastructure-spend patterns while most value flows to a few hyperscalers. The value-capture caution for any firm betting heavily on the current vendor stack, paired with Marcus as the critical-balance limb.

Ju & Aral - Collaborating With AI Agents (RCT)

A preregistered experiment with 2,234 people: human-AI teams produced +50% output per worker but more homogeneous results, and the skills of a good human teammate are not the skills of a good AI collaborator. AI literacy is really a delegation-and-communication skill, and it is trainable.

Percipient - How Frontier AI Models Perform on Real Legal Work

Lawyers with 25+ years' experience blind-graded frontier models: on document review, nine of ten clustered within eight points; on insurance-coverage analysis, the gap reached 37 points. Model choice barely matters for commoditised review and is decisive for complex advisory work, which is also the argument for metering the commoditised tier hard.

- Macro

The State Switches the Model Back On

After nineteen days dark, Fable 5 returned globally on 1 July once the US Commerce Department lifted its export controls; Mythos 5 stays gated to the government-approved Glasswing list. Frontier-model access is now something a government withdraws and restores by directive, which is the sovereignty point from DD1 made operational.

Thomson Reuters - The $143bn Gap

The Future of Professionals survey puts $143bn of US client revenue at risk from the implementation gap, with 78% of corporate clients calling AI quality gains essential and 6% saying providers deliver. The clearest "cost of inaction" and client-pressure data of the fortnight, and a sharp proposal stat.