Paper Instruments Releases Ultramarine 1 Flash AI Model, Launches API

Paper Instruments released its knowledge-work AI model Ultramarine 1 Flash on October 14, 2026. The company also launched an API service for the model and said the model weights would be released two weeks later.

Ultramarine 1 Flash is built on the open-source base model GLM-5.3 Flash and was further trained with large-scale reinforcement learning, according to aitimes.com. Post-training cost about 300,000 US dollars and was carried out by a five-person team.

Training used modified LoRA and CISPO techniques, with a context length of 131,072 tokens and a maximum of 250 agent execution steps. On Paper Instruments' own PI Knowledge Work evaluation, the model scored 59.3 points at 0.13 US dollars per task, compared with 57.8 points at 8.19 dollars for Fable 5 and 55.1 points at 2.09 dollars for GPT-5.6 Sol.

Ultramarine 1 Flash scored higher than its base model on all six knowledge-work benchmarks evaluated, with an average improvement of 4.1 points, while producing 6.5% fewer output tokens including reasoning. It outperformed Fable 5 on four of the six benchmarks, scoring 62.9 versus 57.4 on JobBench and 53.3 versus 49.2 on FrontierFinance, but trailed Fable 5 on DiligenceBench and APEX-Agent 1.1.

API pricing is 0.15 US dollars per million input tokens, 0.03 dollars for cached input and 0.50 dollars for output, the same as the base model. Paper Instruments is also releasing a Paper Office package for safely manipulating office files, along with training code, agent execution infrastructure and grading tools. A desktop app called Feather lets non-developers use the model, and the company said its goal is to replace GPT or Claude in professional agents for banking, law firms, private equity and consulting.