OpenAI now prices the underlying model transparently. The harder economics sit in content rights, review labor and control of the systems where legal work is produced.
OpenAI is moving from supplying models to legal-technology vendors toward competing for a larger share of the legal workflow itself.
Astra for Law, launched September 17, is aimed at large law firms and corporate legal departments. It combines legal research, document analysis and drafting with integrations into existing legal systems. Business Insider reported that OpenAI is targeting major firms, including the Am Law 200, while working with companies such as Harvey, Legora, Thomson Reuters, iManage and DeepJudge.
The launch makes legal AI a useful test of vertical-software economics because the cost of generating a first draft is only one input into a much more expensive production process. Law firms sell reviewed work product whose provenance, confidentiality and professional accountability must survive clients, courts and regulators.
$10/$50 token pricing makes model cost visible; application economics remain opaque
OpenAI publishes GPT-6 Astra API pricing of $10 per million input tokens and $50 per million output tokens. That makes one part of the cost stack unusually measurable. It says much less about who ultimately captures the value created above the model layer.
Legal content rights, matter permissions, document management, citation checking, version control, audit trails and partner review can all cost more than raw inference. Thomson Reuters owns proprietary legal information and the Westlaw distribution channel. Harvey and Legora have built matter interfaces and law-firm relationships around foundation models. iManage controls a large installed base of documents and permissions. OpenAI can compete with those companies while still relying on them for data, distribution or workflow access.
The same ambiguity appears in OpenAI's case study with Legora. The companies said Astra reviewed 41 financial documents in minutes and improved performance on a planted-error benchmark by roughly 40%. That is company-reported evidence, but it points to the right economic unit: verified completion of a legal task, including the review burden that follows the model output.
For a law firm, productivity improves only when lower first-pass labor is greater than the extra validation burden shifted to more expensive lawyers. A contract drafted in half the time creates little operating leverage if a partner spends the saved time finding missing context, repairing clauses or checking citations.
OpenAI also assumes more application obligations as it moves vertically. Legal search requires source provenance. Contract work requires permissioning and version control. Enterprise deployment requires retention policies, security review and integration with the systems where firms already keep privileged material.
That is why Astra matters before its revenue is known. OpenAI is testing whether model leadership can be converted into workflow ownership in a profession where data rights and senior review are unusually expensive. The commercial evidence will come from paid expansion inside large firms, sustained use on consequential matters and falling total review hours. Those outcomes will determine whether the economic surplus accrues mainly to OpenAI, specialist legal-software vendors, proprietary-data owners or the firms themselves.
Sources
- Business Insider, "OpenAI launches Astra for Law": https://www.businessinsider.com/openai-launches-astra-for-law-targeting-legal-tech-industry-2026-9
- OpenAI, "Legora financial statement review with Astra": https://openai.com/index/legora-financial-statement-review-with-astra/
- OpenAI GPT-6 Astra model pricing: https://developers.openai.com/api/docs/models/gpt-6-astra