Sign in Subscribe
3 min read

DeepSeek V4 Pro Matches GPT-5.2 as OpenAI Considers IPO and White House Proposes Vetting

Model Developments & Benchmarks

DeepSeek V4 Pro Matches GPT-5.2 Performance: DeepSeek V4 Pro has successfully matched the performance of GPT-5.2 on the agentic FoodTruck Bench, released just ten weeks after the OpenAI model. Benchmarks indicate that DeepSeek’s version is approximately 17 times cheaper while maintaining high levels of consistency and performance.

Qwen3.6 27B Local Performance Benchmarks: The Qwen3.6 27B FP8 model was demonstrated running on a single consumer-grade RTX 5000 PRO GPU, achieving 80 tokens per second with 200k tokens of context. The setup highlights the increasing efficiency of advanced AI models for local agentic coding applications.

New "Peanut" Text-to-Image Model Debuts: An anonymous text-to-image model called Peanut has reached the #8 spot in the Artificial Analysis Text to Image Arena. The model is expected to surpass popular open-weight models like FLUX.2 when its weights are released in the near future.

Microsoft VibeVoice Ported to C++: A new developer port, vibevoice.cpp, allows Microsoft’s VibeVoice speech-to-speech model to run locally on CPU, CUDA, and Metal without a Python environment. The project supports voice cloning and long-form transcription with diarization across multiple hardware backends.

OpenAI Codex Star Count Spikes: OpenAI's Codex has seen a sudden surge in popularity on GitHub, leading to discussions about its current standing against newer tools like Claude Code. Analysts suggest the spike in interest may be due to operational changes rather than specific technological breakthroughs.

OpenAI Considers Initial Public Offering: OpenAI President Greg Brockman testified during the Elon Musk trial that the company is exploring the possibility of an IPO. The testimony also revealed Brockman's personal stake in the company is valued at approximately $30 billion.

Anthropic Predicts Automated AI Research by 2028: Anthropic co-founder Jack Clark believes AI is nearing a threshold where it can automate its own research and development. He predicts that by the end of 2028, AI will likely be capable of reproducing papers and optimizing its own training protocols independently.

Product Integrations & Tools

Perplexity Computer Expands to Teams and 3D Creation: Perplexity Computer is now available within Microsoft Teams to assist with research and analysis directly in the workspace. Additionally, a new browser tool called Extrudio leverages Perplexity Computer to convert SVGs and pixel art into 3D models without requiring a login.

Mistral AI Utilized for Family Management Apps: A developer has integrated Mistral AI into the Busy Days app to handle OCR, event extraction, and recipe creation for families. The implementation highlights Mistral's performance in parsing structured data from everyday life tasks efficiently.

Policy & Regulation

White House Proposes Vetting Process for AI Releases: The White House is considering new regulations that would require AI models to be vetted before their public release to address cybersecurity concerns. The move has sparked debate regarding government overreach and its potential impact on the pace of AI innovation.