There is a new name buzzing on Hacker News, and it isn't another trillion-parameter behemoth from a Big Tech lab. Enter Ultramarine 1 Flash, a lean, mean, knowledge-work machine from a five-person San Francisco startup called Paper Instruments. While the giants are fighting for general intelligence, Paper Instruments is playing a different game: professional-grade precision at a fraction of the cost.
Specialized for the 'White Collar' Grind
Ultramarine 1 Flash isn't trying to write poetry or code a video game. It's built specifically for the high-stakes world of finance, law, and consulting. The goal is simple: generate the decks, memos, and models that professionals put in front of clients without the eye-watering API costs of frontier models.
According to the lab, the model was built by applying large-scale reinforcement learning to the GLM-5.3 Flash base. The result? A model that reportedly outperforms Anthropic’s Claude Fable 5 on four of six professional-task benchmarks while generating fewer tokens and costing significantly less—some claims suggest up to 60x cheaper.
Claimed vs. Verified
As with any disruptive release, the community is treating the numbers with a healthy dose of skepticism. Critics note that most of the performance data comes directly from Paper Instruments, and public overall scores are still sparse. However, the promise of open weights and a focused API for document agent migration has already caught the attention of developers looking to swap out expensive GPT or Claude workflows.
Whether Ultramarine 1 Flash is a genuine 'Fable-class' killer or just a highly optimized niche tool remains to be seen, but it proves that small teams can still make huge waves in the LLM space.
Sources
Media
