Paper Instruments drops Ultramarine 1 Flash for pennies per task

Ask a frontier model to build an audit-ready financial model or draft a legal memo, and you will usually burn through expensive tokens just trying to fix formatting and tone.

Paper Instruments introduced Ultramarine 1 Flash, an open model built strictly for heavy knowledge work across finance, law, consulting, and corporate operations. The premise is straightforward: deliver client-ready corporate output for pennies.

Built for real client work, not chatbot banter

Ultramarine 1 Flash is fine-tuned on top of the open GLM-5.3 Flash base model, but its outputs look closer to the work of a junior management consultant. The company claims it writes lean prose with strict attention to information hierarchy and visual structure.

Instead of evaluating casual chat, Paper Instruments tested it against tasks that matter to enterprises. It is expected to spit out finished Excel workbooks, formatted PowerPoint decks, and Word documents, while navigating live ERP and MES business software.

API pricing matches its base model: $0.15 per million input tokens, $0.03 for cached inputs, and $0.50 per million output tokens.

Squeezing frontier performance into pennies

According to Paper Instruments, Ultramarine 1 Flash costs an average of roughly $0.13 per task on its internal PI Knowledge Work benchmark, which correlates heavily (r = 0.81) with the Artificial Analysis Intelligence Index.

At thirteen cents per job, the model reportedly outscored Claude Fable 5 ($8.19 per task) and GPT-5.6 Sol ($2.09 per task). It also came within three points of top-tier heavyweights like Claude Opus 5.5 ($6.96) and GPT-6 Sol ($1.26).

On the public JobBench evaluation, it landed a score of 62.9, outpacing rivals like Gemini 3.8 Flash and DeepSeek V4.1 Flash.

Part of this cost efficiency comes from conciseness. Across six core benchmarks, Ultramarine 1 Flash generated 6.5% fewer tokens than the base GLM-5.3 Flash, taking a more direct path to the solution rather than hallucinating corporate filler.

API access is available immediately, with the open weights scheduled to drop soon.

Why it matters

For analysts, legal teams, and boutique consulting shops, this is a very compelling release. Running dozens of complex report generations a day on premier closed models adds up quickly; knocking that cost down to thirteen cents changes the unit economics entirely. Once the open weights drop, privacy-conscious finance firms can run these workflows on their own servers.

For the AI landscape, it highlights an uncomfortable shift for closed frontier labs. Trying to be everything to everyone is expensive, but targeting a specific high-value workflow on top of a cheap open base can beat a giant generalist model on its own turf. At sixty times cheaper than Claude Fable 5 on knowledge tasks, it puts serious pricing pressure on enterprise tiers.

We think Paper Instruments made a smart bet here by focusing on real deliverables rather than generic synthetic reasoning tests. That said, their biggest claims still lean on an internal benchmark. How well it handles dirty enterprise data and legacy ERP quirks in the wild remains to be seen.