GPT-5.6 and Grok 4.5 Launch Amidst Debate on Open Source AI Governance

Key Takeaways

1. OpenAI released GPT-5.6 models (Sol, Terra, Luna) and ChatGPT Work agent, with Sol achieving SOTA performance and significant cost efficiency.
2. Grok 4.5 from xAI leads multiple benchmarks, showcasing strong agentic capabilities and ambitious future plans.
3. Andrew Ng and Yann LeCun warned that US AI governance threatens open source, calling for decentralized innovation.
4. NVIDIA introduced Flex-Forcing video generation and highlighted energy as the ultimate AI bottleneck.

1. OpenAI Launches GPT-5.6 and ChatGPT Work

1. OpenAI officially launched the GPT-5.6 model family (Sol, Terra, Luna) and a new agent, ChatGPT Work, powered by Codex, capable of cross-application file operations and long-term project tracking. — via 1 2
2. Sol is the SOTA model, excelling in programming, knowledge work, science, and cybersecurity while using fewer tokens and lower costs. Sol Ultra achieves ~750 TPS in computer use. — via 1 2
3. GPT-5.6 Sol becomes the first frontier model to beat the ARC-AGI-3 game, scoring 7.8% SOTA, and has been integrated as the preferred model for Microsoft 365 Copilot. — via 1 2
4. Ethan Mollick validated GPT-5.6's real-world utility by using it to build a game, proofread a book (30 min, dozens of notes, no errors), but criticized agent product design for non-programmers. — via 1 2 3

2. Grok 4.5 Leads Benchmarks with Strong Agentic Capabilities

1. Elon Musk announced Grok 4.5 achieves top scores on AutomationBench-AA, Terminal-Bench v2, Harvey Legal, SWE Marathon, and more, with superior cost efficiency. — via 1 2
2. Grok 4.5 demonstrated powerful agentic and coding abilities: creating a FPS game in one hour, building 3D scenes in UE5.8, and constructing mathematical counterexamples. The Grok Build tool continues to improve. — via 1 2
3. Future plans include expanding context to 1M tokens, a new 2T parameter base model completing this month, and monthly foundation model releases. — via 1

3. Open Source AI Under Threat: Calls for Decentralized Innovation

1. Andrew Ng criticized US AI governance for imposing informal review systems that effectively force compliance on open source developers, warning that a "restrict until allowed" approach could kill open source models. He drew parallels to the 1990s Clipper Chip battle. — via 1
2. Yann LeCun asserted that open source is the bedrock of US tech dominance and all efforts to restrict open source in AI must be resisted, cautioning against power concentration in a few proprietary AI providers. — via 1 2

4. Other Notable Developments

1. NVIDIA released Flex-Forcing, a video generation method that allows switching between bidirectional diffusion and autoregressive methods at inference. They also released quantized Qwen3.6 with 2.5x speedup on GPUs. — via 1 2
2. Jensen Huang compared AI infrastructure to a five-layer cake with energy at the bottom, calling current capex the largest in history but constrained by grid capacity. — via 1
3. Hugging Face completed the Gemma challenge (100+ agents achieved 5x speedup on Gemma 4 inference) and the Build Small hackathon (946 apps, 817 developers, 22 winners). — via 1 2