OpenAI Forced To Delay GPT-5

OpenAI CEO Sam Altman has delayed the launch of GPT-5 by several months, citing integration challenges and unexpectedly high performance gains uncovered during development. As a substitute, OpenAI will release the o3 and o4-mini models in the approaching weeks—previously planned as internal components of GPT-5. The o3 model, specifically, has been internally benchmarked at the extent of a top-tier coder.

Altman pointed to a few key reasons for the shift: the complexity of merging features right into a single system, the necessity to scale infrastructure for unprecedented demand, and the chance to push GPT-5 far beyond initial expectations. The o-series will function interim steps, offering scalable architecture, multimodal capabilities, and cost-effective inference.

In a recent legal development, a federal judge denied OpenAI’s motion to dismiss, calling its argument a “straw man” and upholding the NYT’s contributory infringement claim. The court noted evidence that OpenAI knew its models could reproduce copyrighted content. A brand new study also suggests that OpenAI’s models, including GPT-4 and GPT-3.5, can have “memorized” copyrighted material like books and news articles during training. OpenAI maintains its training practices align with fair use, however the ruling strengthens the NYT’s legal challenge.

On the business front, OpenAI reportedly discussed buying Jony Ive and Sam Altman’s AI device startup in line with The Information. Read more.

Meta has launched two models from its Llama 4 family: Scout and Maverick, each now integrated into Meta AI across WhatsApp, Messenger, and Instagram.

Maverickdesigned for general assistant tasks, comprises 400 billion total parameters but prompts only 17 billion across 128 experts. Scoutoptimized for summarization and code reasoning, supports a ten million token context window and might run on a single Nvidia H100 GPU.

Meta also previewed Llama 4 Behemothstill in training, with 288 billion energetic parameters, positioning it as a future top-tier base model and reportedly outperforms models like GPT-4.5 and Claude 3.7 Sonnet on STEM benchmarks. Nevertheless, none of those models qualify as “reasoning” models that fact-check answers.

CEO Mark Zuckerberg added that a fourth model, Llama 4 Reasoning, can be introduced inside weeks. All models employ a mixture-of-experts (MoE) architecture to spice up efficiency and multimodal performance. Despite their open-weight release, Meta has barred EU-based developers from using the models, citing regulatory uncertainty under the EU AI Act.

The Maverick model, ranked second on LM Arena, has raised concerns on account of discrepancies between its benchmarked and public versions. Researchers noted that the LM Arena version is an experimental, “chat-optimized” model, fine-tuned for conversation, while the publicly available model differs, using excessive emojis and long-winded answers. This raises serious questions on the integrity of benchmarks, suggesting some models are engineered to perform well in tests while leaving developers in the dead of night about their true real-world capabilities. Read more.