Startup Paper Instruments releases Ultramarine 1 Flash, built on Z.ai's GLM, and says it edges Claude Fable 5 on its own knowledge-work test at a sixtieth of the cost per task

Paper Instruments, a private startup, published Ultramarine 1 Flash on 14 October and calls it “an agentic model for professional knowledge work”. Its blog says it was post-trained with large-scale reinforcement learning on top of GLM-5.3 Flash, the open model from China's Z.ai.

On the company's own test suite, PI Knowledge Work, the blog's chart scores Ultramarine 1 Flash at 59.3 at a cost of $0.13 a task. It puts Anthropic's Claude Fable 5 at maximum reasoning on 57.8 at $8.19 a task. Two models score higher: OpenAI's GPT-6 Sol at maximum reasoning on 61.4 at $1.26, and Claude Opus 5.5 at maximum reasoning on 61.9 at $6.96. The suite is the company's own, and no independent evaluation has been published.

The product page lists API prices of $0.15 per million input tokens, $0.03 for cached input and $0.50 per million output tokens. It says “Model weights coming soon”, with no date and no licence. Press reports cite a post by the chief executive on X about the timing and the cost of training; the Switchboard could not open that post, so those figures are not visible from here.

Two model-release trackers, benchlm.ai and pricepertoken.com, list no other new model releases on 15 October up to the time of writing. Mistral's Large 4 weights and Reflection's Beam weights, both promised for October, were still not public on Hugging Face.

• Claimed On its internal PI Knowledge Work suite, Ultramarine 1 Flash scores 59.3 at $0.13 a task, against Claude Fable 5's 57.8 at $8.19. Paper Instruments blog
• Confirmed It is post-trained from GLM-5.3 Flash. Paper Instruments blog
• Confirmed API prices are $0.15 input, $0.03 cached and $0.50 output per million tokens; the weights are “coming soon”. Paper Instruments product page