The Sequence Radar #885: Last Week in AI: Models, Games, and the Future of Evaluation

There's a pattern this week that's hard to miss once you see it: every major AI move was really about action, not conversation. OpenAI's limited preview of GPT-5.6 arrived with three variants named Sol, Terra, and Luna — a flagship, a balanced option, and a cheap, fast one. That's not just branding. It signals that companies no longer want one giant model; they want intelligence tuned to different jobs, from frontier reasoning down to high-volume automation. More telling than the benchmarks was the rollout itself, wrapped in a safety framework and phased access, government coordination included. Releasing a model is starting to look less like shipping software and more like deploying regulated infrastructure.

Anthropic, meanwhile, quietly shipped something smaller but arguably more revealing: Claude Tag, which lets users mark prompts and responses with explicit semantic labels. It sounds minor. But as models take on longer, more autonomous tasks, loose conversational prompting stops being enough — you need structure, roles, and context that persist. Claude Tag is a small bet that the future of prompting looks more like designing a workflow than writing a clever sentence.

Then there's General Intuition, the Medal spinout that just raised $320 million at a $2.3 billion valuation to mine something nobody else is chasing at scale: action-labeled gameplay. Their argument is that a video clip of someone playing Fortnite or Minecraft isn't just footage — it's a record of perception, decision, and consequence. That's exactly the kind of data current models lack when they try to reason about the physical world. If the open web trained language models, gameplay data might train embodied ones. It's a strange, compelling bet, and investors clearly think so too.

And then evaluation got genuinely fun. The LayerLens Stratix Cup pitted sixteen models against each other playing soccer, with Claude Opus 4.8 beating GPT-5.5 1-0 in the final. Forget the scoreline — the real story is that these models had to write their own strategies, adapt mid-game, and operate under imperfect information. That's a much harder test than answering a multiple-choice benchmark. It's behavior under pressure, not prose under review.

Taken together, these four stories point to the same shift: models are being asked to sense, plan, and act inside messy environments, not just produce good answers to fixed questions. The competitive edge won't just be about who has the biggest model anymore. It'll be about who builds the best sandboxes, the tightest guardrails, and the most revealing games to find out what these systems can really do when nobody's holding their hand.