OpenAI did confirm that it had begun training a new “frontier model” on May 28, 2024. It also announced a board-level Safety and Security Committee with a 90-day timetable for reviewing safeguards. But OpenAI did not call the system GPT-5, publish a launch date, or say the model could not ship before the review ended. The “at least 90 days” claim was a contemporary interpretation of the governance schedule, not a release promise.
This is a historical fact-check of that announcement, not a current forecast of OpenAI product availability.
What OpenAI actually announced on May 28, 2024
OpenAI’s announcement combined two developments. First, the company said it had “recently begun training its next frontier model.” Second, it said the board had formed a Safety and Security Committee to make recommendations on critical safety and security decisions across OpenAI’s projects and operations.
The company described the committee as responsible for evaluating and further developing OpenAI’s processes and safeguards during an initial 90-day period. It would then submit recommendations to the full board, after which OpenAI said it would publicly share an update on recommendations adopted, subject to safety and security considerations. The announcement is available from OpenAI.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →OpenAI also used aspirational language about reaching “the next level of capabilities” and advancing toward artificial general intelligence. Those statements were the company’s characterization of its ambitions, not an independently verified measurement of the model’s capabilities.
Was GPT-5 officially confirmed?
No. The announcement never used the name GPT-5. It referred only to a “next frontier model.” Journalists and industry observers inferred that the successor to GPT-4-class systems would probably receive the GPT-5 name, but that was shorthand and expectation rather than an official designation.
| Question | What the announcement established |
|---|---|
| Was a new model in development? | Yes. OpenAI said training had recently begun. |
| Was it called GPT-5? | No name was provided. |
| Was a launch date announced? | No. |
| Were capabilities, modalities or pricing disclosed? | No. |
| Was a 90-day release delay promised? | No. The 90 days applied to the committee’s review process. |
What the 90-day period covered
The 90-day clock concerned the committee’s work on safeguards and processes, not an explicit countdown to public access. Counting 90 days from May 28, 2024 produces approximately August 26, 2024, but that was an expected point for recommendations to reach the board—not a guaranteed launch date.
The announcement did not say that the committee could block a release, that the model could not launch before its recommendations, or that the model would launch immediately afterward. Board review and product-release decisions could have followed separate schedules.
Why the interpretation seemed plausible
A substantially more capable model would reasonably be expected to undergo additional evaluation, security work and deployment planning. Because OpenAI announced the training cycle and the safety review together, contemporary coverage treated the review as a likely constraint on timing. VentureBeat’s headline made that inference explicit: “GPT-5 won’t come for at least 90 days.”
Why it was still only an inference
The public statement did not identify the review as the only launch condition or establish a direct dependency between committee recommendations and model availability. OpenAI could have released an intermediate system, offered a staged deployment, changed the product name, continued evaluating the model, or chosen not to release it publicly at all.
Who was on the Safety and Security Committee?
The initial committee was led by board chair Bret Taylor and directors Adam D’Angelo, Nicole Seligman and Sam Altman. OpenAI also listed technical and policy participants Aleksander Madry, Lilian Weng, John Schulman, Matt Knight and Jakub Pachocki.
The company said it would retain or consult additional specialists, including former cybersecurity official Rob Joyce and former Justice Department official John Carlin. On June 18, 2024, OpenAI updated the announcement to note that retired U.S. Army General Paul M. Nakasone had joined the board and the committee.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Was the committee independent?
Independence was a central governance question, not a settled factual conclusion. The body consisted mainly of OpenAI directors, an executive and internal technical or policy personnel. Critics argued that a group reviewing the company’s own safeguards did not provide the same independence as an external regulator or a genuinely independent audit panel.
OpenAI’s response, as described in its announcement, was to consult outside safety, security and technical experts. Consultation does not give those experts decision-making authority, so it should not be described as independent oversight unless OpenAI separately grants that authority.
What the review could have affected
OpenAI published a broad remit rather than a detailed checklist. Areas plausibly relevant to the committee’s work included:
- release safeguards and deployment restrictions;
- preparedness and risk evaluations;
- cybersecurity protections for model weights and infrastructure;
- escalation and board-approval procedures;
- post-release monitoring and incident response.
These are reasonable areas of relevance, not a published list of guaranteed deliverables or evidence that any particular control was adopted.
Why “training has begun” did not mean launch was near
Starting training says little about a release candidate’s readiness. A frontier-model program can involve several distinct stages:
- Pretraining: building the base model from large-scale data and compute.
- Post-training: improving instruction following, behavior and usefulness.
- Safety evaluation and red-teaming: testing misuse, dangerous capabilities, robustness and security risks.
- Productization: preparing inference capacity, interfaces, monitoring, documentation and access controls.
- Deployment approval: deciding whether and how to release the system, potentially through limited or staged access.
The May announcement did not reveal how far pretraining had progressed, whether post-training had started, whether internal capability targets had been met, or how long evaluation would take. “Begun training” therefore cannot establish that a public launch was imminent.
What readers could conclude at the time
| Claim | Status |
|---|---|
| A successor-level frontier model was being trained | Confirmed by OpenAI |
| The system was officially named GPT-5 | Unconfirmed |
| A 90-day safety-process review was planned | Confirmed by OpenAI |
| GPT-5 could not launch for at least 90 days | Inference from the review timetable |
| A launch date or product specification existed publicly | Not announced |
Why the announcement mattered beyond a possible product name
The statement tied a new model-training cycle to board-level governance at a time when OpenAI faced intense scrutiny over safety practices, departures and control of powerful AI systems. Its significance was therefore institutional as much as technical: OpenAI was presenting safety and security review as part of how it would govern frontier development.
That framing still left practical questions unanswered. The announcement did not specify the committee’s veto power, publish measurable acceptance criteria, or explain exactly how its recommendations would govern deployment. It also placed an internal committee at the center of oversight while relying on external experts in a consultative role.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesHow to interpret the headline today
The headline was directionally based on a real announcement but compressed several uncertainties into a product prediction. The defensible statement is that OpenAI had started training an unnamed frontier model and scheduled a 90-day review of safety processes. Calling that model GPT-5 and converting the review deadline into a minimum launch delay went beyond what OpenAI had confirmed.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
