Who can stop OpenAI from releasing a model?

A signed agreement with California’s attorney general sets out who can halt an OpenAI model release. The power sits with a safety committee of the OpenAI Foundation. The committee can require safeguards or stop a release even if OpenAI’s risk thresholds would otherwise allow it. The agreement was signed on 27 October 2025, the day before the company announced its recapitalisation. Executed California memorandum

OpenAI’s 2023 board crisis shows how hard it can be to make a decision stick. Directors removed Sam Altman and appointed an interim chief executive. Altman returned before the month was out. The later agreement gives the Foundation’s Safety and Security Committee its own authority over model releases. Whether the committee has used that authority to halt a release remains unresolved.

At its launch in December 2015, OpenAI described a nonprofit that would pursue artificial intelligence for broad public benefit without needing to generate a financial return. Its later Charter addressed how to share benefits and avoid uses of AI that would unduly concentrate power. Those are different statements made at different stages of the organisation. 2015 launch announcement · OpenAI Charter

In 2019 OpenAI created a capped-profit structure to attract capital. Its announcement said returns for first-round investors were capped at 100 times their investment, with individual caps negotiated in advance and lower caps expected for later rounds. That year Microsoft announced a $1 billion investment and became OpenAI’s exclusive cloud provider at that stage. The partnership gave OpenAI access to computing capacity and gave Microsoft a relationship with a lab building new models. OpenAI’s 2019 structure announcement · 2019 Microsoft partnership announcement

On 17 November 2023 the board said it had lost confidence in Altman and questioned his candour with directors. It did not give specific examples. On 29 November OpenAI confirmed Altman’s return and named Bret Taylor, Larry Summers and Adam D’Angelo as its initial board. Altman thanked Microsoft for offering to take in the team. The company’s announcements establish the public sequence. They leave the relative influence of employees, directors and Microsoft Unresolved. Removal announcement · Return announcement

Former director Helen Toner later alleged that the board learned of ChatGPT’s release on Twitter and that Altman had failed to disclose his control of OpenAI’s Startup Fund. In the same episode, TED carried OpenAI chair Bret Taylor’s response, which cited the WilmerHale review commissioned after the crisis. OpenAI’s published summary said a breakdown in trust precipitated the removal and that Altman’s conduct did not mandate it. Taylor’s response and the summary did not address Toner’s examples individually. The full review and underlying board material were unavailable. The specific allegations remain Contested. Toner interview and Taylor’s response · OpenAI’s published review summary

In October 2025 the operating company became a public benefit corporation. OpenAI disclosed a 26% Foundation equity stake at the close. The current equity percentage remains Unresolved. The legal control rights operate through a special share. While the Foundation holds it, the Foundation can appoint or remove operating-company directors. Its written approval is required for specified changes, including changes to the mission and governance rules. OpenAI’s structure disclosure · California memorandum

The safety committee’s authority belongs to the Foundation, and the agreement provides for access to the operating company’s staff and research. These are Verified written powers. Some directors also serve on both boards. The sources reviewed here do not establish whether the committee has halted a release or how directors would act if they disagreed with the operating company. California memorandum · Delaware’s statement of non-objection

OpenAI said on 1 September that it delayed parts of Astra’s development and release while it strengthened and tested safeguards. The announcement does not identify the Foundation or its Safety and Security Committee as the decision-maker. Whether the committee used its formal authority remains unresolved. OpenAI’s Astra announcement, 1 September 2026

OpenAI says the new structure helps it raise the money and hire the people needed to pursue its mission. The scale is visible. In March 2026 the company announced a financing close of $122 billion in committed capital at a post-money valuation of $852 billion. The release leaves individual payment schedules undisclosed. Some named investors also sell infrastructure or related services. The disclosures do not establish what proportion of OpenAI’s capital came from suppliers overall or whether a particular investor’s money paid its own invoices. OpenAI’s structure disclosure · OpenAI’s March financing announcement

Amazon’s filing for the quarter ended 30 September reports that it completed its $50 billion OpenAI Series C investment, including $21.3 billion in the third quarter. The filing also describes a separate AWS agreement. It does not trace investment proceeds to AWS invoices. SoftBank said on 1 October that it completed its $30 billion follow-on investment, bringing its reported cumulative investment to $64.6 billion and ownership to approximately 13% at the time of disclosure. Those investor-specific figures do not establish total proceeds across the round, later ownership or further financing. Amazon September 2026 Form 10-Q, Notes 1–2 · SoftBank follow-on investment announcement, 1 October 2026

The April 2026 Microsoft amendment shows what commercial influence can look like without a board appointment. The companies say Microsoft remains OpenAI’s primary cloud partner, with first release on Azure subject to the stated exception, and OpenAI can serve customers through other providers. Microsoft keeps a non-exclusive licence to models and products through 2032. OpenAI continues to pay Microsoft a share of revenue through 2030 under a total cap whose amount was not disclosed. The Foundation’s written safety authority can coexist with those commercial obligations. Whether the obligations influence an individual release decision requires evidence about that decision. The companies’ April 2026 announcement

In Musk v. Altman, the district court adopted the advisory jury’s finding that the claims tried in May 2026 were brought too late. That finding settled the timing issue for those claims. In July a separate order appointed a mediator while other matters remained. The orders examined here do not decide whether OpenAI kept its founding promise. The current disposition of the remaining claims and any appeal remain Unresolved. May order, ECF 580 · July order, ECF 606

Evidence cutoff 11 November 2026. This account compares original company announcements, the signed California memorandum, Delaware’s statement, published responses to the 2023 dispute and the May and July court orders. Company accounts of private deliberations are attributed. The full WilmerHale review, current ownership figures and a complete current docket were unavailable.

The Foundation kept substantial formal authority after the recapitalisation, including the safety committee’s power to halt releases. Commercial pressures are also visible in the company’s financing and infrastructure agreements. Evidence of the committee using its authority, a current ownership disclosure, or a fuller record of the 2023 board decision would help establish how these pressures affect decisions in practice. The public record gives readers a way to ask the question when the next controversial model is ready to ship.