LegalTech: Power players

GenAI has dominated headlines in the wake of renewed tensions between frontier AI models and the US government which are only partially resolved. On 13 June, US authorities ordered Anthropic to suspend foreign nationals, including Anthropic’s own employees, from using its most capable products, Claude Mythos 5 and its public counterpart Claude Fable 5, because of national security concerns. These likely centred on its ability to identify bugs and vulnerabilities in critical software. In order to comply with this, Anthropic announced that it was disabling Fable 5 and Mythos 5 for all its customers.  

On 26 June, the US government allowed Anthropic to re-release Mythos 5 to a limited list of organisations responsible for critical infrastructure in the US. ‘We received notice from the US government that Mythos 5, our strongest cybersecurity model, can be redeployed to a small group of cyber defenders and infrastructure providers,’ Anthropic said in a statement. However, the rollback did not include permission to re-release Fable 5. 

Meanwhile, OpenAI announced that it was limiting the rollout of its new GPT-5.6 models to a ‘small group of trusted partners’ that have been cleared by the US government. The Information reported that both Anthropic and OpenAI have blunted or planned to blunt their models’ cybersecurity capabilities when releasing them to the broader public.

Switching off public access to frontier AI models reignited the debate around sovereign AI, which is preoccupying governments, organisations and cybersecurity professionals. 

The reaction was not limited to UK and European politicians calling for urgent investment in sovereign AI. The Wall Street Journal reported on Friday that US government roadblocks to US systems are incentivising the use of Chinese AI systems which claim to match the performance of Mythos in some cybersecurity scenarios. GLM-5.2, a new open-source model released by China’s Zhipu AI, also known as Z.ai, can match US models when it comes to finding security bugs in software. And because it is open source, it is much cheaper to use than Anthropic and OpenAI models, even though it is less powerful. Chinese cybersecurity company 360 Security Technology claims that its new tool, Tulongfeng, is comparable to Mythos in finding bugs. Speaking at a cybersecurity conference in Beijing, 360 Security chief executive Zhou Hongyi said: ‘This kind of powerful weapon that can alter the landscape of cyberwarfare can’t remain solely in American hands.’

Legal AI sovereignty

The legal sector has been exploring the concept of sovereign AI on its own terms. More firms are developing proprietary AI models, mostly partnering with general-purpose infrastructure models Anthropic, OpenAI, Palantir and Microsoft (which makes sense given that most law firms are Microsoft customers).

A newer trend is for legal AI vendors to develop proprietary AI models. Thomson Reuters announced early access for existing customers to the latest version of CoCounsel, rebuilt on Anthropic’s Claude Agent SDK (software development kit), which offers law firms ‘fiduciary-grade AI’. At the same time, they are making progress with a proprietary large language model, Thomson, currently being tested at Imperial College London, which may operate in tandem with CoCounsel and other Thomson Reuters products. 

Meanwhile, Harvey is working with several research partners to develop and fine-tune large language models to help law firms build and own proprietary legal AI models (rather than partnering with one of the US general-purpose models currently being limited by government orders).

Vibe-coding at LegalTechTalk

Vibe-coding, which produced open-source platforms MikeOSS and Lavern, is a way for firms of all sizes to maintain independence from vendors and their pricing models. While vibe-coded solutions are not always scalable, they are a great way to build prototypes and point solutions while retaining control over the AI’s capabilities and limitations – and firm and client data.

In its second year, LegalTechTalk has grown into the legal sector equivalent of CES (the Consumer Electronics Show) in Las Vegas, with multiple stages, sessions and exhibition halls, and 5,500 attendees. Unsurprisingly, there was a strong focus on AI. The main takeaway is that legal is no longer a comfortable fast follower, but in the unpredictable vanguard of AI transformation.

One new experience was the Vibeathon, organised by vibecode.law founders Matt Pollins of Lupl, Alex Baker of the Legal Tech Collective and Chris Bridges of Tacit Legal. Vibe-coding involves building apps on an agentic AI platform using natural language prompts. To participate, all you had to do was show up with your laptop and an idea, and the team would set you up on Replit, an agentic AI platform that writes production-ready code and helps you turn your idea into a useable app. This was my first experience of vibe-coding, and (with guidance from the team and law firm volunteers) I built an editorial assistant to proofread articles, flag up production checks and apply my personal style guide. While clean development depends on giving the AI clear instructions, there are many details to consider. 

Vibe-coding lets you build a workable app quickly. But you cannot see the path the coding agent chose, which makes iteration challenging. And each iteration uses tokens, so the cost of compute depends on the complexity of the app and how efficient your prompting is. 

As Bridges wrote on LinkedIn, while vibe-coding does not produce enterprise-grade tools, it adds value because it ‘collapses the time taken to prototype and test an idea from months to days’. 

There are also enterprise vibe-coding solutions which do produce enterprise-grade tools, but they are not free or open source. Fliplet, originally a no-code drag-and-drop app builder for web and mobile, has evolved into an AI-first enterprise vibe-coding solution. Natural language prompts are used to create apps, and there are options to bring in extra features such as integrations with existing tools, databases and workflows, automate processes, manage access and check security. A big advantage is the ability to track version history, which means you can revert to a previous iteration. Building in analytics enables you to track usage rates. 

CEO Ian Broom explains that Fliplet enables law firms and other regulated sectors to vibe-code apps that comply with professional information security standards. Fliplet’s enterprise-grade vibe-coding tools are a bit like Thomson Reuters’ ‘fiduciary-grade AI’ models. As Fliplet’s website states: ‘The hard part is not only generating the first version. It is deciding how the software will be governed, secured, published, maintained and trusted once real users depend on it.’  

Agentic usage-based pricing 

'What we are really watching is the end of the AI subsidy era reaching the legal market. For two years, the true cost of inference sat hidden behind venture-funded flat rates'

Hélder Santos, Bird & Bird

Agentic AI signalled the end of the cheap AI era. CEO Winston Weinberg told Sourcery that when Harvey switched from chat-based products to cloud agents, its monthly token consumption increased from 1 trillion in January to between 12 and 13 trillion in May. In a separate episode, Harvey co-founder and president Gabe Pereyra drew an analogy between the billable hour and usage-based pricing. He envisaged that the accountability mechanism that customers will eventually demand will look like the legal six-minute increment: per token visibility with auditing and routing layers built on top. Pereyra added that vertical companies (like Harvey) are positioned to provide that layer because they can show ROI (return on investment) per token.

Bird & Bird’s global head of legal technology and innovation Hélder Santos highlights the shift from per-seat to consumption pricing as a function of the maturing legal AI market. ‘What we are really watching is the end of the AI subsidy era reaching the legal market. For two years, the true cost of inference sat hidden behind venture-funded flat rates, and agentic tools burn far too much compute for that to hold. Usage-based pricing is just the honest version of what was always coming,’ he says. 

Last week, Legora introduced consumption-based pricing for its Agent Pro product, presenting the change as a shift from licences to outcomes. But as legal tech commentator Elgar Weijtmans wrote on LinkedIn, ‘consumption is not the same as outcome’. However, because ‘each run can be tied to the matter that prompted it… [consumption-based pricing] finally makes the cost of AI legible client by client in a way the seat never could’. 

This echoes Pereyra’s analogy. Santos observes: ‘The irony is that legal AI was meant to kill time-based billing, yet the vendors are building a metered, usage-based model of their own. Not [based on] hours, but units. The disruptor adopting the model of the disrupted.’