DeepLearning.AI Weekly: GPT-5.6 Family Limited Release, Sakana Model Orchestration, and Microsoft's In-House Reasoning Model

The DeepLearning.AI weekly roundup covers several developments: OpenAI released three closed-source vision-language models, GPT-5.6 Sol, Terra, and Luna, currently available only to about 20 U.S. government-approved organizations, with prices ranging from $5 to $1 per million input tokens; OpenAI says it will expand access in the coming weeks. Sakana AI released Fugu and Fugu-Ultra, two models that orchestrate other models and agents, claiming they surpass Claude Fable 5, Claude Mythos Preview, and GPT-5.5 on benchmarks such as Terminal-Bench 2.1 and GPQA-Diamond. Microsoft launched MAI-Thinking-1, its first reasoning model not distilled or fine-tuned from another vendor, using a MoE architecture with 1 trillion total parameters and 35 billion activated per token, claiming it surpasses Claude Sonnet 4.6 with 97.0% on AIME 2025. The roundup also introduced RoboReward, a robot reward model from teams at Stanford and Berkeley.