Skip to content
Neural Daily

ModelsNews

Ultramarine 1 Flash: knowledge work for 13 cents a task

Paper Instruments unveiled Ultramarine 1 Flash, an agentic model tuned for professional knowledge work. It builds on the open GLM-5.3 Flash model and, on the company's internal benchmark, costs $0.13 per task while outscoring Claude Fable 5.

By: Mikola Máté Patrik · · 2 min read

Paper Instruments graphic: blue brushstroke above the words Ultramarine 1 Flash
Image: Paper Instruments · Press kit · Official launch image · Source

What it is

Ultramarine starts from the open GLM-5.3 Flash base model, which Paper Instruments further trained with large-scale reinforcement learning on non-verifiable rewards (Paper Instruments). The curriculum spans finance, law, consulting and operations, with 19 practice areas and 88 environment types. Tasks require finished Excel, PowerPoint and Word deliverables or work inside live business systems such as ERP.

The numbers

The company says Ultramarine beats its base on all six knowledge-work benchmarks it tested, by 4.1 points on average, while generating 6.5% fewer tokens. On its PI Knowledge Work suite it costs $0.13 per task, outscoring Claude Fable 5 ($8.19) and GPT-5.6 Sol ($2.09) and landing within three points of Claude Opus 5.5 ($6.96) and GPT-6 Sol ($1.26), all at max reasoning (Paper Instruments).

Important caveat: PI Knowledge Work is an internal benchmark. The company says it correlates with the Artificial Analysis Intelligence Index at r = 0.81, but there is no independent evaluation yet.

Pricing and access

The API is OpenAI-compatible and also works with the Anthropic SDK. Pricing is $0.15 per million input tokens, $0.03 cached input and $0.50 output, the same as GLM-5.3 Flash. Weights are listed as coming soon on Hugging Face. The company also released packages based on python-docx, python-pptx and openpyxl for Word, PowerPoint and Excel work.

Why it matters

If the numbers hold up independently, it is more evidence that targeted RL can turn a cheap open base model into a near-frontier professional model. For office agents and report generation, it is worth benchmarking against a pricier frontier model on your own tasks.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles