Paper Instruments' Ultramarine 1 Flash targets office work at cents per task

Ultramarine applies large-scale reinforcement learning to GLM-5.3 Flash, an open base model[^1]. Paper Instruments calls the approach non-verifiable RL: instead of training only on tasks with a single checkable answer, it rewards the model against rubrics for substance, accuracy and presentation on finished work such as Excel models, PowerPoint decks and Word documents[^1][^2]. The company says this gives the model stronger prose and visual hierarchy than typical models, and that prompts need tuning to get the most from it[^3].

Paper Instruments reports that Ultramarine improves on its GLM-5.3 Flash base across all six knowledge-work benchmarks it tested, by 4.1 points on average, while generating 6.5% fewer tokens[^1]. On its internal PI Knowledge Work suite it scores 59.3 at $0.13 per task, ahead of Claude Fable 5 at 57.8 ($8.19) and GPT-5.6 Sol at 55.1 ($2.09), and within three points of Claude Opus 5.5 and GPT-6 Sol[^1]. The company says those internal scores correlate with the Artificial Analysis Intelligence Index at r = 0.81[^1].

The public benchmarks are more mixed. Ultramarine tops JobBench at 62.9 and FrontierFinance at 53.3 among the models Paper Instruments compares, but trails Gemini 3.8 Flash and Claude Fable 5 on DiligenceBench and APEX-Agents 1.1[^2]. All of these are the company's own runs.

The interesting part is not the leaderboard but the price. If a small post-trained model really produces client-ready spreadsheets and memos at around a sixtieth of the cost of a frontier model, the economics of document-heavy agents change: you can afford to generate several drafts and have a stronger model judge them. Treat the vendor scores as a starting hypothesis. Run it on ten of your own real deliverables, have someone who does the work grade the output, and wait for the weights before you plan on running it in-house.