Paper Instruments releases Ultramarine 1 Flash for RL on non-verifiable tasks

Paper Instruments has unveiled Ultramarine 1 Flash, an open model built on GLM-5.3 Flash and trained with reinforcement learning on professional knowledge work in finance, law and consulting. It reaches Fable-class benchmark scores at $0.13 per task — 60x cheaper than Claude Fable 5 ($8.19) and within three points of Claude Opus 5.5.

• Training spans 88 environment types across 19 professional practice areas
• $0.13 per task versus $8.19 for Claude Fable 5
• Within three points of Claude Opus 5.5 on knowledge work benchmarks
• Rewards come from task-rubric co-design, not verifiable answers