Ultramarine 1 Flash: What if five researchers were enough to compete with Claude?

Paper Instruments unveils Ultramarine 1 Flash, an AI model specialized in professional work, trained for $300,000 and announced to be 60 times cheaper.

Producing a financial model, preparing a presentation for an executive committee, or analyzing a legal file requires more than a well-written response. It requires understanding documents, complying with industry constraints, and delivering an actionable result. Ultramarine 1 Flash, developed by Paper Instruments, focuses specifically on these tasks. Its team claims performance comparable to Claude Fable 5, for an operating cost up to 60 times lower in its evaluations. Even more surprising: its post-training reportedly required only five people and around $300,000. A specialized model rather than one trained from scratch

Ultramarine 1 Flash is built on GLM-5.3 Flash, an open model developed by Z.ai. Paper Instruments did not train a new foundation architecture from scratch, but instead enhanced the capabilities of an existing base using a reinforcement learning method geared toward knowledge work professions.

The training targets four domains: finance, law, consulting, and operations. Tasks range from producing Word documents and PowerPoint presentations to building Excel models, as well as interacting with professional software such as ERPs, CRMs, and industrial management systems.

The objective is to train an AI capable of producing directly usable deliverables, rather than simply explaining how to create them. Learning to judge work without a single right answer

The distinctiveness of Ultramarine lies in its training process. In programming, a model can be evaluated based on tests that pass or fail. But how do you automatically determine whether a strategic memo is relevant, whether a presentation adheres to consulting firm standards, or whether a financial analysis correctly distinguishes facts from assumptions?

Paper Instruments developed a system that generates synthetic professional scenarios, accompanied by evaluation criteria tailored to each task.

This system covers 19 professional specialties and 88 environment types, split across finance, law, consulting, and operations. The scenarios combine documents, communications, regulatory constraints, and industry-specific software.

Each deliverable is evaluated on precise criteria, focusing notably on information accuracy, consistency of conclusions, and compliance with instructions. This approach makes it possible to train the model on work whose quality cannot be determined by a simple comparison with an expected answer. Performance close to Claude for a fraction of the price

On JobBench, Ultramarine 1 Flash reaches 62.9 points, compared to 57.4 for Claude Fable 5 and 60.6 for Gemini 3.8 Flash. It also scores 53.3 points on FrontierFinance, versus 49.2 for Claude Fable 5.

Results are more mixed on other evaluations. Ultramarine lags behind Claude Fable 5 on DiligenceBench and APEX-Agents 1.1, two benchmarks evaluating more complex professional tasks.

The economic comparison is particularly highlighted on PI Knowledge Work, an internal evaluation by Paper Instruments. Ultramarine scores 59.3 points there at an average cost of $0.13 per task, compared to 57.8 points and $8.19 for Claude Fable 5, in its maximum reasoning configuration.

The difference represents a cost approximately 63 times lower within this specific protocol. However, these comparisons come from a set of evaluations combining public results, tests run by Paper Instruments, and internal benchmarks. They do not demonstrate a general superiority of the model. Feather, an environment designed for real work

To train Ultramarine, Paper Instruments also developed Feather, an execution environment for AI agents.

It allows the model to view files, edit documents, execute commands, and interact with external tools. Operations are structured to preserve project state and prevent, for example, an agent from overwriting a file modified in the meantime.

Paper Instruments accompanies this setup with Paper Office, a suite of software libraries tailored for manipulating Word, Excel, and PowerPoint documents.

The company also indicates that it worked on writing behavior and deliverable presentation, specifically visual hierarchy, phrasing clarity, and adherence to professional standards. An available API, open weights announced

Ultramarine 1 Flash is accessible via API, priced at $0.15 per million input tokens, $0.03 for cached inputs, and $0.50 per million output tokens.

The model can also be tested in Feather, the professional environment developed by Paper Instruments. Access is being rolled out gradually based on available capacity.

The company plans to release the model weights and its training code within two weeks following the announcement. Their release is therefore not yet effective at launch.

Ultramarine 1 Flash illustrates another path in the AI model race: starting from an open base and investing in specialized learning rather than systematically seeking to build a larger model. This strategy could interest companies developing agents for consulting, finance, or law professions, where the quality of the deliverable and its production cost matter more than the general performance of a conversational assistant.