Search, costs and safety

Open-source corpus mapping and enterprise cost controls lead the edition, while OpenAI’s Astra delay underlines cyber-safety constraints on deployment.

Briefs

Models and research

• OpenOntology targets cheaper agentic search — DOSS plans to open-source OpenOntology, which maps large corpora for agentic search. Its evaluation says Haiku 4.5 with the system matched Opus 5 on measured tasks while using less context, time and money.

• NVIDIA explores transferable KV caches — NVIDIA researchers describe a way to convert a model’s KV cache for use by another model, so the target can skip processing the same context again. The technical explainer says conversion can run 2.7 to 25 times faster than reprocessing, which could speed systems that switch models mid-task.

Enterprise AI

• Databricks sets out AI coding cost controls — Databricks argues that firms should compare models by task quality and cost, then route work to the least expensive capable option. Its announcement and technical write-up also recommend spend visibility, progressive limits and smaller contexts. Databricks says these methods can cut some unit costs by up to 90%.

• AI adoption can conceal weak productivity — Varick argues that a small group of power users often gains most of the value from enterprise AI, while many staff barely use it. Its analysis and article recommend tracking work that is manual, hybrid or automated, rather than log-ins, and embedding agents in existing tools.

Safety and deployment

• OpenAI delays Astra over cyber safety — Sam Altman says OpenAI will take longer to make Astra generally available because of the model’s cyber capabilities. The release update shows that cyber-safety work can delay deployment of frontier models.