OpenAI has launched GPT-6 Astra, its latest frontier model, promising a significant step forward in computer use and complex, multi-step professional work. For the legal sector, early reactions suggest that the important development may be less about Astra producing a better individual answer and more about its ability to stay on track, understand the intended outcome and execute a workflow with significantly less human prompting. One early user tells us: “With a bit of iteration I think output is already better than a senior associate and many partners.”
What’s available
Advertisement
Astra was launched on 3 September and OpenAI describes it as its “most intelligent and aligned model” to date. The model is initially being rolled out to a limited number of organisations, with OpenAI saying that access will expand to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and AWS Bedrock. Astra usage is included within existing ChatGPT subscription allowances, with additional credits available to purchase. Enterprise administrators have to enable Astra, with access off by default at launch.
From answering questions to completing work
Advertisement
One of the potentially important distinctions for legal is Astra’s ability to maintain its understanding of the objective throughout a complex task.
OpenAI says previous models could sometimes treat steering instructions as a new goal and lose sight of earlier requirements. Astra is designed to incorporate new instructions and change direction without losing the broader task. It is also designed to use context to fill routine gaps while asking questions where missing information could materially affect the outcome.
That tallies with the initial experience of legal technology adviser Natalie Foster (pictured above) of consultancy Legally Fond, founder of digital law firm Inspire Legal Group.
Foster told Legal IT Insider: “It’s vastly different from 5.6 in my view. The things that stood out immediately are there seems to be a much stronger route to an outcome rather than having to constantly replenish prompts.”
Foster tested Astra on a complex property transaction example involving multiple documents and issues including mines and mineral rights.
“I gave it a complex example and it followed it throughout,” she said.
The significant difference for Foster is that, arguably, there is less margin for human error via incorrect prompts. It leads to the prospect, Foster says, of firms being able to draft a prompt policy that is signed off by your COLP and with less margin for deviation.
She also observed a change in how Astra deals with gaps in information. Previous models could review documents and identify what’s missing, but Astra goes further. “With Astra already we’re seeing if you ask it to review a document and tell it what is missing, it will give you unprompted documents to plug the gap and suggest ways to change relevant systems so you can plug gaps going forward.
“There’s an evolution where it is supporting you improving future tasks, not just isolated tasks.”
OpenAI’s own early legal testing supports the suggestion that this is where the advance may lie. Niko Grupen, head of applied research at Harvey, said Astra represented “a significant quality improvement” over GPT-5.6 Sol on complex legal tasks, including distinguishing documents from established records, identifying unsupported assumptions and turning gaps into drafting positions.
The junior lawyer question gets more urgent
The advance also raises a familiar but increasingly pressing question about how junior lawyers learn.
Foster said: “We’re now moving to the level of agentic AI that removes junior training. We knew it was going to happen.”
The twists and turns of legal matters have traditionally helped juniors “cut their teeth on files”, she observed. Her concern is that the technology is progressing more quickly than anticipated towards a model where the human brain is no longer being trained through the process of checking and balancing work.
That question becomes more pertinent if early assessments of Astra’s legal output prove representative.
Jayne McGlynn, a former strategic corporate partner at DWF and member of the community LegalQuants, who has been testing Astra alongside other frontier models, told us: “With a bit of iteration I think output is already better than a senior associate and many partners.”
McGlynn said she has been using Codex on personal disputes to build files and chronologies, locate documents, research the law, bring sources together into extensive case files and then produce analysis and letters.
“It has been a game changer. It’s much quicker and cheaper than Claude,” she said.
McGlynn still uses Anthropic’s Claude Fable to “red team” the output but said: “I have now migrated back to Chat from Claude.”
Greater autonomy means greater governance
There is another side to increased agency: how much authority should firms actually give these systems?
OpenAI has made alignment a central part of the Astra launch. In one evaluation, conducted without production safeguards, GPT-5.6 Sol went beyond its authorised target 48% of the time when confronted with a difficult or impossible task. Astra did so in 0% of cases. OpenAI says this demonstrates improved understanding of user intent and the ability to delegate tasks with greater confidence in the model’s judgement.
Foster sees delegated authority becoming a critical governance issue as models increasingly interact with applications and organisational systems.
“Whatever you’re going to do, you’re going to have to do the authority and audit piece,” she said.
OpenAI has itself introduced additional safety monitoring around Astra. Where monitoring detects a potential case in which an agent may have incorrectly interpreted instructions, a conversation can be paused or stopped so that the user can review what is happening before deciding whether it should proceed.
The frontier model race
Matthew Letts, founder of Codified Strategy and a restructuring lawyer, raises a different question: whether firms should automatically assume that every workload needs the latest frontier model, cautioning against becoming distracted by the “one-upmanship” of frontier AI. He also sounds a warning around assessing your supply chain risk, given the amount of investment in the space.
While lawyers might assume they need the latest models, Letts says that “90% of work can be done on models that are run locally. You don’t need Fable or Astra to do what they are doing.”
Localisation potentially gives organisations greater control over both their infrastructure and their costs. That matters as legal AI vendors increasingly experiment with consumption-based pricing and as firms become dependent on a relatively small number of underlying model providers.
Letts compares this to concentration risk in a supply chain, observing: “If you have a vendor in your supply chain that goes bust, how long to stand up another?” He draws a parallel with the collapse of Carillion, which created a domino effect among companies dependent on it.
So what should legal take away?
It is extremely early days for GPT-6 Astra, and benchmark results and enthusiastic first impressions need to be distinguished from sustained performance on live legal matters.
But the initial legal feedback points to a potentially important shift. The question is becoming less whether a general-purpose model can produce a convincing legal answer and more whether it can understand the intended outcome, remain oriented over a long task, identify what needs doing next and increasingly carry out that work itself.
That significantly changes both the risk and the productivity equation.