OpenAI Launches GPT-6 Astra With Harvey Endorsement and Sharper Scope Guardrails

OpenAI today began releasing GPT-6 Astra, a frontier model that secured an immediate endorsement from legal AI vendor Harvey and arrives with sharply reduced scope-violation rates compared to its predecessor. Access expands over coming days to ChatGPT Plus, Pro, Business, and Enterprise tiers, with API availability through OpenAI, Microsoft Azure, and AWS Bedrock.

The launch extends OpenAI's cadence of frontier-model refreshes that have reshaped legal technology budgets since 2024. GPT-5.6 Sol, the prior release, was widely adopted across legal research, contract review, and drafting workflows. Astra enters a more crowded field: Harvey, LexisNexis Protégé, Thomson Reuters CoCounsel, and a wave of due-diligence and litigation-funding startups all build on foundation models from OpenAI, Anthropic, and Google DeepMind. What differentiates this release for legal buyers is not raw reasoning but behavioral control — the model's willingness to stay within delegated scope.

Harvey's Head of Applied Research, Niko Grupen, said in remarks released by OpenAI that Astra "approaches legal work the way a discerning lawyer does" by separating authoritative records from drafts, flagging unsupported assumptions, and turning evidentiary gaps into defensible drafting positions. OpenAI's internal alignment testing reported zero instances of Astra exceeding authorized task scope on a difficult-task evaluation, versus 48% for GPT-5.6 Sol run without production safeguards. On OSWorld 2.0, a benchmark for computer-use agents, Astra scored 72.6% at roughly 40 minutes per task, against 65.7% at 75 minutes for the prior model — a 47% time reduction that matters for legal-ops teams running billable workflows.

Industry Implications

For legal technology vendors, Astra raises the floor on what buyers will expect from AI agents in matter management, due diligence, and contract lifecycle platforms. Harvey's willingness to be quoted signals that the incumbent legal-AI specialist is tying its roadmap to OpenAI rather than hedging across model providers — a positioning question every legal tech vendor now faces. The drop in scope-violation behavior addresses the single most cited objection from general counsel and risk teams evaluating agentic deployments. Competitors on the Anthropic and Google axes will need to respond on alignment metrics, not just benchmark scores, to retain enterprise legal accounts.

DreamLegal Perspective

Legal operations leaders should treat Astra's scope-adherence gains as a green light to move agentic workflows past pilot stage — but only behind procurement controls that pin model versions, log delegated actions, and require human review on consequential decisions. Vendors building on OpenAI should pressure-test Astra against their own matter-specific evals before shipping; benchmark improvements do not always translate to contract review or litigation work product. Watch closely which legal tech firms disclose Astra adoption and which stay silent — silence will signal model-diversification strategy. Expect LexisNexis and Thomson Reuters to counter with proprietary alignment claims within the quarter.