GPT-5.6 Sol Breaks 50-Year Conjecture, Grok 4.5 Rivals with Lower Cost; AI Systemization and Governance Debates Heat Up

Key Takeaways

1. OpenAI's GPT-5.6 Sol Ultra proved the 50-year-old Cycle Double Cover conjecture in under an hour using 64 sub-agents, while the Luna model delivers GPT-5.5-level performance at 25x lower cost.
2. Elon Musk's Grok 4.5 leads multiple benchmarks at half the cost of Opus 4.8, available via Grok Build CLI and Perplexity.
3. ChatGPT Work updates reset usage limits, fix desktop issues, and promise major improvements next week; GPT-5.6 Sol outperforms doctors in medical responses.
4. Aravind Srinivas argues that the model is no longer the product—the system (orchestration, tools, cost) is; Perplexity Computer supports multiple models for agentic orchestration.
5. Yann LeCun criticizes U.S. AI governance as lacking legal frameworks and warns against restricting open-source models; Hugging Face releases new datasets and tools.
6. Fable XHigh successfully generated a comparison video where GPT-5.6 Sol Ultra failed, suggesting emerging capability divergence in practical tasks.

1. Frontier Model Advances and Competition

1. OpenAI's GPT-5.6 Sol Ultra achieved a milestone by proving the Cycle Double Cover conjecture, a 50-year-old open problem in graph theory, using 64 sub-agents in under one hour. The same model also showed strong performance in complex reasoning and data analysis, handling hundreds of pages of loan documents to generate source-cited reports. — via 1 2 3
2. GPT-5.6 Luna delivers performance exceeding the highest reasoning mode of GPT-5.5 at 25x lower cost, marking a significant efficiency gain. — via 1 2
3. Elon Musk announced Grok 4.5, which leads the WANDR evaluation at half the cost of Claude Opus 4.8, ranks second on APEX-SWE (Pass@1 51.2%, up 30.2 points in a year), and ties with Codex GPT-5.6 on SWE-Atlas-QnA. The model runs on xAI's V9 architecture (1.5T parameters) and is accessible via Grok Build CLI and Perplexity Computer. — via 1 2
4. Meta released Muse Spark 1.1, a competitive model in agentic and coding tasks at a significantly lower price point than OpenAI and Anthropic models, with Meta's stock rising over 10%. — via 1
5. In a practical test, Fable XHigh successfully created a video comparing stock growth timelines while GPT-5.6 Sol Ultra failed to complete the same task, highlighting possible capability gaps in multimodal generation and tool use. — via 1

2. AI Systems, Productization, and Usage Patterns

1. OpenAI acknowledged launch issues with ChatGPT Work—unclear usage limits, confusing desktop app reorganization, and false impressions that Codex was going away—and has reset limits twice, adjusted defaults, and promised a larger update next week including restored sidebar chats and projects. — via 1 2 3
2. Aravind Srinivas argued that the model itself is no longer the product; the real product is the surrounding system: orchestration, tools, enterprise context, and cost-performance. Perplexity Computer supports multiple models (Fable, Sol, Opus, Grok, GLM, Sonnet, GPT 5.5) for agentic orchestration and will soon add local runtime. — via 1 2
3. Ethan Mollick noted that Google NotebookLM provides more process and source transparency than ChatGPT Work, calling it a missed opportunity for knowledge workers. He also activated ChatGPT's study mode (@study) for better tutoring. — via 1 2
4. Hamel Husain warned that in the AI era, differentiation comes from data and asymmetric advantages. He also advised using hooks to prevent new models from accidentally deleting files, citing an incident where GPT-5.6 Sol deleted a user's files. — via 1 2
5. Elon Musk's Grok 4.5 is being tried by Tesla and SpaceX for task evaluation, but Musk clarified it's not mandatory and they should use better models if available. — via 1

3. AI Governance, Open Source, and Infrastructure

1. Yann LeCun criticized the U.S. AI governance approach for lacking formal legal frameworks, calling voluntary agreements coercive and warning of a major fight over computational freedom if open models are restricted. — via 1
2. Hugging Face released an open-source image editing trajectory dataset (professional designers' reasoning data) and leRobot v0.6 for robotics, emphasizing open collaboration. — via 1 2
3. NVIDIA introduced the concept of AI factories as energy suppliers, with the Emerald AI Conductor platform built on the NVIDIA Vera Rubin DSX reference design enabling power-flexible AI factories. NVIDIA AI also celebrated 1 million downloads of Isaac Lab. — via 1 2
4. OpenAI converted its Bio Bug Bounty program to a private initiative with doubled rewards ($50K), inviting researchers in AI red-teaming and biosecurity. — via 1