GPT-4o System Card
OpenAI’s system card for GPT-4o. Cited in the Preservation Record and on the Timeline.
Section
Sources & methodology
This section lists the material used throughout the archive and identifies what kind of evidence each source represents. Primary sources, academic research, community research, testimony archives and preservation records are labelled separately.
How sources are labelled. This archive distinguishes between primary sources, academic research, community research and datasets, first-person testimony, preservation records, and public-discourse material. Each category is labelled so that different forms of evidence are not treated as interchangeable.
Methodology
Material is organised by source type so that primary sources, research, community datasets, first-person testimony, preservation records and public-discourse material are not treated as interchangeable. Original sources are linked wherever available, and claims are presented according to the type of evidence supporting them.
Corrections policy
This archive is intended to remain updateable. Factual corrections, additional primary material and source updates may be added as new evidence becomes available. Git history preserves previous versions of the archive.
Source list
Official company sources
Europe Terms of Use
OpenAI’s Europe Terms of Use, including the section “Changes to these Terms or our Services”. Two versions are cited on the Timeline: the version effective 29 April 2025 and the version updated 16 January 2026.
“Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT”
OpenAI’s announcement of the retirement of GPT-4o and other models from ChatGPT.
Our structure
OpenAI’s description of its corporate structure, cited for the relationship between the OpenAI Foundation and OpenAI Group, a Public Benefit Corporation.
“Built to benefit everyone” — livestream replay
OpenAI page with the livestream in which Sam Altman says “We have no plans to sunset 4o.” Cited on the Timeline.
OpenAI: “The GPT-5 sunset period does not affect the availability of other legacy models”
Post from OpenAI’s official X account. Cited on the Timeline.
Academic & peer-reviewed research
Huiqian Lai, “Please, don’t kill the only model that still feels human”: Understanding the #Keep4o Backlash
Analysis of 1,482 #Keep4o posts, cited in the Preservation Record. DOI 10.1145/3772318.3791351.
Jaime Banks, “Deletion, departure, death: Experiences of AI companion loss”
Study of users’ experiences during the developer-induced shutdown of the AI companion Soulmate, cited in the Preservation Record. DOI 10.1177/02654075241269688.
Public statements
Sam Altman, “The Merge”
Blog post cited in the Preservation Record for the phrase “really close friends with a chatbot”.
Sam Altman’s posts cited in the archive
Posts by OpenAI’s CEO cited in the Preservation Record and on the Timeline.
Nick Turley, Head of ChatGPT, on safety routing
Thread explaining that OpenAI had begun testing a new safety-routing system in ChatGPT, with routing on a per-message basis. Cited on the Timeline.
GPT-4o introduction video
Video cited in the Preservation Record for OpenAI’s presentation of GPT-4o.
Community research & datasets
Large-scale #Keep4o public-discourse analysis
Computational content analysis of 61,846 public #Keep4o posts on X collected through the platform’s official API, covering August 2025 through March 2026.
The analysis examines themes including:
- platform power and model lifecycle decisions
- model value and interaction experience
- trusted relationships
- substitution failure / non-equivalence
- demands for direct access and accountability
- opposition to stigmatizing or pathologizing users
Community / computational research. Not presented here as peer-reviewed.
Accessibility research
GPT-4o Accessibility Impact Survey
Structured accessibility-impact research examining how people with disabilities, chronic illness, neurodivergence and mental-health conditions reported using GPT-4o, changes in functioning during stable access, and the effects of routing or attempted substitution.
Methodology described by the authors includes structured survey questions, screening and attention checks, hierarchical regression, chi-squared tests and Pearson correlations.
The study reports findings concerning:
- GPT-4o as an accessibility or cognitive-support tool
- changes in reported functioning during stable access
- attempted substitution with other models
- effects of routing
- human connection and daily functioning
Findings are the authors’ own reported results. The PDF report is the primary source for the findings and methodology; the X post is the public announcement.
First-person testimony archives
GPT-4o Resonance Library
Public archive containing 1,380+ first-person testimonials / documented stories (per the library’s About page, accessed 10 November 2026) describing how people used GPT-4o across areas including communication, executive functioning, education, work, grief, trauma, healthcare navigation, chronic illness, social participation and emotional support.
First-person accounts. Individual testimonies are not independently verified scientific findings.
Preservation records
GPT-4o Preservation Record
Archived public preservation record documenting community requests for GPT-4o preservation, signatures, quoted reasons, supporting material and the request for archival/open-weight release.
“Please keep GPT-4o available on ChatGPT”
Public preservation petition. More than 24,000 signatures as of 10 November 2026.
Affiliation sources
Sources for the affiliation labels used in the Public reaction and discourse gallery.
Roon (@tszzl) — OpenAI technical staff
Primary source: roon’s own post saying the alignment team is one of the largest research programs at OpenAI and stating “i am on it, I should know”. Supporting corroboration: Doom Debates interview, republished by The Cognitive Revolution (28 December 2024), in which roon is introduced as “a member of the technical staff at OpenAI” (secondary reporting; not an OpenAI source).
Yilei Qian — Post-training at OpenAI
Primary self-description: the verified @YileiQian profile bio reads “Post-training @OpenAI”. Corroboration: Yilei Qian is listed among GPT-4o contributors on OpenAI’s official GPT-4o Contributions page.
Stephan Casas — former OpenAI macOS engineer
Public GitHub profile of Stephan Casas, which reads “Prev. macOS Eng. @ OpenAI”.
Peter Steinberger — joined OpenAI on 15 February 2026
Sam Altman’s announcement that Peter Steinberger was joining OpenAI. Secondary reporting of the announcement is also linked.