This week in AI — cheaper frontier models, design that peels into layers, and video that remembers the cast

This was not a one-logo week. Frontier got cheaper to run at volume, design tools got more editable, and video research started treating continuity like a real production problem. Four signals worth keeping on the desk.

**Embedded evaluation.** On Sep 21 Anthropic and Accenture put third-party evaluators inside enterprise Claude rollouts. Outsiders in the room while the work is still being shaped, not only after the model ships. Access got a staffing answer. Independence of who pays is still an open craft question.

**Volume-tier frontier.** Sep 23 brought Claude Opus 5.5 and GPT-6 Sol, with Luna as the high-volume sibling. Near-flagship capability at much lower token prices, with adjustable reasoning effort. Ten days after the pace-the-frontier essay, the market answered with a production bill of materials. Capability stayed high. The invoice got friendlier.

**Design that peels.** Sep 22–23 Ant Group’s inclusionAI open-sourced Ming-Image-0.1-Design and a companion Layer model under MIT. Generate a text-rich layout, then peel it into editable transparent layers a human can still move. Flat generation is common. Getting the composition back into layers is closer to how branding and web teams actually iterate.

**Video that remembers the cast.** On Sep 24 Google Research published an AI video co-director on Gemini and Veo. Multi-agent orchestration for multi-shot continuity, with companion frameworks for visual memory and closed-loop fixes. Research headed to academic venues, not a consumer button yet. The craft signal is continuity labor getting a co-pilot so someone can keep directing.

Builders still own taste. The tools are catching up to the boring hard parts: evaluation in the room, cheaper volume, editable layers, and characters that stop mutating between cuts.

Primaries: https://www.anthropic.com/news/accenture-embedded-evaluation · https://www.anthropic.com/claude-opus-5-5 · https://huggingface.co/inclusionAI/Ming-Image-0.1-Design · https://research.google/blog/coherent-long-form-video-generation/