Ultramarine 1 Flash Rewrites Knowledge-Work Costs

The real story is specialization: frontier-level economics may increasingly come from focused post-training and better task environments, not ever-larger general-purpose models.

• –Ultramarine leads three of six knowledge-work benchmarks, but places third on both DiligenceBench and APEX-Agents, so “Fable-class” is workload-dependent.
• –Its strongest results target finished Excel, PowerPoint, and Word deliverables; it is less dominant on open-ended agent loops and automation.
• –Synthetic tasks, custom rubrics, and domain-specific environments produced a 4.1-point average gain over the base model while using 6.5% fewer tokens.
• –The $0.13 versus $8.19 comparison uses Paper Instruments’ internal suite and official API pricing, making it compelling but not independent proof of general frontier parity.
• –API access is live, but model weights are still forthcoming; self-hosting and outside replication will determine how durable the advantage is.