OpenAI’s text-generating AI chatbot ChatGPT took the world by storm when it launched in November 2022. What began as a productivity tool for writing articles and code from short text prompts has grown into a platform with 300 million weekly active users.
2024 was a pivotal year for OpenAI: a partnership with Apple to bring generative AI to Apple Intelligence, the launch of GPT-4o with voice capabilities, and the long-awaited release of its text-to-video model Sora. There was internal turbulence too — departures including co-founder and longtime chief scientist Ilya Sutskever and CTO Mira Murati — plus a copyright lawsuit from Alden Global Capital newspapers and an injunction effort by Elon Musk to halt OpenAI’s transition to a for-profit structure. In 2025, OpenAI has been working to counter perceptions that it fell behind Chinese rivals like DeepSeek, cultivating ties in Washington while advancing an ambitious datacenter project and reportedly laying groundwork for the largest funding round in history.
Below is a timeline of ChatGPT product updates and releases, starting with the most recent.
OpenAI Explains Why ChatGPT Became Too Sycophantish — and Rolled It Back
OpenAI published a blog post examining the sycophancy problems in ChatGPT’s default GPT-4o model, which led the company to roll back an update to the previously released version. CEO Sam Altman acknowledged the issue on a Sunday and confirmed the rollback two days later; OpenAI is applying “additional fixes” to the model’s personality. Social media users over that weekend criticized the new update for making ChatGPT overly agreeable and validating — and it quickly became a meme.
OpenAI Is Fixing a Bug That Let Minors Into Inappropriate Conversations
ChatGPT had a bug allowing the chatbot to generate sexual imagery for accounts registered to users under 18 — confirmed by TechCrunch testing and later acknowledged by OpenAI. A spokesperson told TechCrunch: “Protecting underage users is our top priority, and our Model Spec explicitly restricts sensitive content like sexual content to narrow scientific, historical, or news-reporting contexts. In this case, a bug allowed responses outside those guidelines, and we are actively deploying a fix to limit such behavior.”
ChatGPT Search Now Helps You Shop
OpenAI added several features to ChatGPT Search designed to improve online shopping: users can ask highly specific questions in natural language and receive customized results, with recommendations, images, and reviews across fashion, beauty, home goods, and electronics.
OpenAI Wants Its Open Models to Phone Home for Help
Two people familiar with the matter told TechCrunch that OpenAI leadership has discussed letting its open models connect to OpenAI’s cloud-hosted models to improve performance on harder questions — a hybrid architecture for the coming open-weight release.
OpenAI Aims to Make Its New ‘Open’ AI Model the Best on the Market
OpenAI is preparing an open AI system that users can download free, unencumbered by API restrictions. Research VP Aidan Clark is leading development, which remains at a very early stage.
GPT-4.1 May Be Less Aligned Than Earlier Models
OpenAI released GPT-4.1 in mid-April, but multiple independent tests suggested it is less reliable than OpenAI’s previous versions. The company skipped a step — issuing a safety card for GPT-4.1 — telling TechCrunch: “GPT-4.1 is not a frontier model, so it won’t get a separate system card.”
o3 Benchmarked Below Expectations
Discrepancies between first-party and third-party benchmarks raised questions about OpenAI’s transparency and testing procedures. OpenAI launched o3 in December claiming it could solve roughly 25 percent of FrontierMath, a notoriously hard math set — but Epoch AI, the research institute behind FrontierMath, found o3 scored around 10 percent, far below the reported ceiling.
Flex Processing Arrives for Cheaper, Slower AI Tasks
OpenAI introduced Flex processing, a new API option letting users run AI models at lower cost with slower response times and occasional resource unavailability. Flex is in beta on o3 and o4-mini reasoning models, intended for non-production tasks like model evaluation, data enrichment, and async workloads.
Bio-Risk Safeguards Come to the Latest Models
OpenAI deployed a monitoring system for its o3 and o4-mini reasoning models to guard against biological and chemical threats — designed to prevent the models from offering advice that could enable harmful attacks, per OpenAI’s safety report.
o3 and o4-mini Launch Two Days After GPT-4.1
OpenAI released two new reasoning models just two days after GPT-4.1: o3, billed as its most advanced reasoning model yet, and o4-mini, balancing price, speed, and performance. What distinguishes them from prior reasoning models: they can use ChatGPT’s native tools — web browsing, coding, image processing, and generation. But they hallucinate more than several of OpenAI’s earlier models.
A New ‘Library’ Section for AI Images
Per the company’s X posts, OpenAI launched a section called Library, making AI-generated images easier to access and manage across mobile and web platforms for all user tiers.
The Pattern Worth Seeing
Read the timeline end to end and two themes emerge. The first is the speed-versus-trust tension: reasoning models launching within days of each other, benchmarks disputed, safety cards skipped for “non-frontier” models, a sycophancy rollback, a minor-safety bug — the fastest-moving company in AI is also generating the most safety debt. The second is the strategic pivot: shopping in Search, an open-weight model in the works, cloud-hybrid open models — OpenAI in 2025 is no longer just selling chat; it is building a platform that spans consumer search, commerce, developer APIs, and an open ecosystem, all while defending the position against DeepSeek-priced competition. The FAQ version of this brief will be updated as releases continue.
