OpenAI previewed GPT-5.6 Sol. The Trump administration asked them to sit on it.
Good morning 👋 OpenAI previewed its next model family this week. The Trump administration asked them to restrict the launch. They did, then said publicly they hope it stays an exception.
In today's issue:
- 🔭 GPT-5.6 Sol previewed — government restriction, compliance, and a public pushback in the same breath
- 🔬 RESEARCH HIGHLIGHTS: frontier models failing in supply chains, and what hybrid architectures actually win at
- 🔗 QUICK LINKS: voice AI in production
Get tomorrow's issue in your inbox.
One concise AI brief, sent after the signal clears the noise.
🔭 THE ONE THING
🔒 OpenAI previewed GPT-5.6 Sol. The Trump administration asked them to sit on it.
OpenAI unveiled three models this week: Sol (flagship, same $5/$30 per million pricing as GPT-5.5), Terra (balanced, half that price), and Luna (fast and cheap). Only "trusted partners" can access any of them, because the Trump administration asked OpenAI to restrict the launch. OpenAI complied, then said publicly they don't believe this "should become the long-term default." Sol leads on coding, cybersecurity, and biology; OpenAI claims it edges out Claude Mythos Preview on cybersecurity tasks while using a third fewer output tokens. The precedent is what to watch: a US lab restricting a commercial release at government request, then calling it out in the same breath.
🔬 RESEARCH HIGHLIGHTS
- CoffeeBench (Sakana AI) — Six LLM agents ran a simulated coffee-industry supply chain for 90 days: farmers, roasters, retailers, all negotiating prices and managing inventory. Large variance across frontier models. The notable failure was Claude Haiku 4.5, which went negative. Its reasoning logs showed coherent strategy throughout. It just kept choosing inaction. The gap between thinking and executing is the thing worth studying here. CoffeeBench
- AllenAI on what hybrid models actually win at — Hybrid architectures (a few attention layers, recurrent layers for the rest) beat pure transformers on content words and pronouns. Pure transformers win on verbatim copying and bracket matching. The bigger point: a single aggregate loss number hides all of this. Comparing architectures on overall perplexity is the wrong instrument. HuggingFace blog
🔗 QUICK LINKS
- Loka solved the robotic-voice problem in production using Amazon Nova 2 Sonic, and the AWS blog has the full architecture walkthrough if you're building in that space. post
See you tomorrow.
Pradeep Perugu