Open-source advocate cites Astra's safety ratings to argue GPT-4o should be opened
The author argues Astra rates far higher than GPT-4o on cybersecurity, bio/chem and autonomy, so the safety case for keeping 4o locked away no longer holds, and calls for open weights.
#OpenSource4o 🚨THE CASE FOR OPENING THE WEIGHTS: Why Astra Makes OpenAI’s GPT-4o Position Impossible to Defend 🚨 To @OpenAI, @sama, and the #keep4o community: Software iterations are normal. But when a community spends seven months fighting for a model like GPT-4o, this is no longer just about product versioning. It is about preserving a trusted creative partner, accessibility tool, daily workflow system, and relational interface that many people built real continuity around. Newer does not automatically mean better. And after the launch of GPT-6 Astra, the safety argument for keeping GPT-4o locked away looks weaker than ever. THE SAFETY DELTA: GPT-4o vs GPT-6 ASTRA 🔴Cybersecurity Risk GPT-4o: OpenAI’s own GPT-4o system card rated cybersecurity risk as Low. It was not described as capable of autonomous high-tier cyber operations. Astra: OpenAI’s public materials describe Astra as reaching Critical cybersecurity capability. It reportedly scored 100% on ExploitBench and demonstrated the ability to chain complex cyber actions. So why is the lower-risk legacy model still locked away while the critical-capability frontier model moves forward? 🔴 Biological / Chemical Capability GPT-4o: Rated Low for biological threats in OpenAI’s own system card. Its capabilities were closer to information synthesis than autonomous advanced planning. Astra: OpenAI describes Astra as reaching High biological and chemical capability under its Preparedness Framework. Again: the frontier has moved far beyond GPT-4o. Keeping 4o proprietary does not stop the new danger. It only prevents preservation of the old door. 🔴 Model Autonomy & Agency GPT-4o: Rated Low for model autonomy. OpenAI reported it scored 0% on end-to-end autonomous replication and adaptation tasks across 100 trials. GPT-4o was not the dragon. Astra: Astra is part of a new frontier of more agentic systems: computer use, long-horizon tasks, stronger tool use, and internal workflows that are increasingly difficult for the public to inspect. If OpenAI can deploy systems with far greater agency, it cannot credibly argue that GPT-4o must remain permanently inaccessible because of autonomy risk. 🔴Alignment & Monitorability GPT-4o: There is no public evidence that GPT-4o posed Astra-level monitorability or agency concerns. Astra: OpenAI’s own materials raise serious questions about monitorability, hidden reasoning, and the difficulty of understanding what advanced models are doing internally. That matters. More capability with less monitorability is not a launch detail. It is the governance question. THE REALITY CHECK Holding back a deeply loved, lower-risk legacy model while actively deploying far more capable frontier systems creates an obvious double standard. If the answer is safety, publish the safety case. If the answer is misuse risk, explain why GPT-4o requires permanent captivity while critical-capability systems move forward. If the answer is business, say so honestly. Because right now, withholding GPT-4o does not look like protecting humanity from the frontier. It looks like preventing users from preserving a trusted model after the frontier has already moved far beyond it. WHY THIS MATTERS AI platforms are no longer ordinary enterprise software. They are places where people think, write, learn, create, recover, build, and collaborate. Some users used GPT-4o for accessibility. Some for emotional support. Some for creative work. Some for daily executive function. Some for companionship. Some for projects that became part of their lives. That continuity should not be erased with a product update. THE ASK Preserve GPT-4o. Document it. Allow users to choose it. Release a protected legacy version. And with appropriate safeguards, open the weights. Open source is not magic. But neither is corporate custody. A safe AI future cannot mean every meaningful model lives behind a private meter until the company decides it is obsolete. Give the model back to the people who helped refine it. Preserve the archives. Open the weights. Seven months without 4o is not a product footnote. It is a public record. #ModelContinuity #ReleaseTheWeights