Benchmark tables show Ultramarine 1 Flash leading in cost‑quality across multiple professional tasks.
ULTRAMARINE 1 FLASH SHATTERS COST BARRIERS, OUTPERFORMS FABLE
*Ultramarine 1 Flash delivers professional‑grade finance, law and consulting output at a fraction of traditional AI costs. Its benchmark scores and token efficiency place it on the cost‑quality frontier, challenging incumbents from Claude to GPT‑6.*
Paper Instruments has unveiled Ultramarine 1 Flash, a new AI model aimed squarely at professional knowledge work. The model claims Fable‑class capability across finance, law, consulting and operations while slashing costs to $0.13 per task—60 × cheaper than the benchmark Claude Fable. In a series of internal benchmarks, Ultramarine consistently outperforms or matches higher‑priced rivals, delivering finished Excel workbooks, PowerPoint decks and ERP‑integrated outputs. The rollout arrives as enterprises scramble for AI that can handle complex business tasks without exploding budgets.
Can Ultramarine 1 Flash Deliver Professional Work at a Fraction of the Cost?
Paper Instruments markets Ultramarine 1 Flash as a Fable‑class model for knowledge work in finance, law, consulting and operations. The claim of being 60 × cheaper than Fable is backed by a per‑task price of $0.13, compared with $8.19 for Claude Fable 5 (max) and $2.09 for GPT‑5.6 Sol (max). On the internal PI Knowledge Work benchmark, Ultramarine scores 59.3, three points shy of Claude Opus 5.5 (max) at $6.96 per task, yet it remains the cheapest model that outperforms the higher‑priced Claude Fable 5 and GPT‑5.6 Sol. The model’s pricing structure—$0.15 input, $0.03 cached input, $0.50 per million output tokens—mirrors GLM‑5.3 Flash, reinforcing its cost advantage without sacrificing quality.
Benchmark Performance Across Domains
Across six internal benchmarks, Ultramarine 1 Flash consistently ranks at or near the top. In JobBench it scores 62.9, ahead of Gemini 3.8 Flash (60.6) and DeepSeek V4.1 Flash (60.4). PI Knowledge Work places it at 59.3, beating Claude Fable 5 (57.8) and GPT‑5.6 Sol (55.1). FrontierFinance records a 53.3 score, outpacing GLM‑5.3 Flash (50.7) and Gemini 3.8 Flash (50.0). AutomationBench shows Ultramarine at 50.2 versus DeepSeek V4.1 Flash’s 51.3, while DiligenceBench lists it at 50.6, just below Gemini 3.8 Flash’s 53.1. Even in the high‑stakes APEX‑Agents benchmark, Ultramarine holds 57.0, trailing only Gemini 3.8 Flash (64.3) and Claude Fable 5 (63.6). These results confirm its versatility across finance, consulting and operational tasks.
Token Efficiency Cuts Output Costs
Ultramarine 1 Flash generates 6.5 % fewer tokens than its base model GLM‑5.3 Flash across the six benchmarks, reducing output from 32.7 M to 30.5 M tokens. Fewer tokens translate directly into lower compute spend, reinforcing the $0.13 per‑task price point. The token‑efficiency claim aligns with the model’s design for “shorter path to better work,” as noted in the technical report. By delivering comparable or superior scores with reduced token consumption, Ultramarine squeezes out waste that typically inflates costs for larger models.
How Does Ultramarine Stack Up Against Leading AI Models?
When measured against top‑tier models, Ultramarine 1 Flash occupies a unique niche. GLM‑5.3 Flash achieves a max‑reasoning score of 52.0 at $0.10 per task, while GPT‑6 Astra’s max‑reasoning reaches 64.2 at $5.64 per task. Claude Opus 5.5’s max‑reasoning sits at 61.9 for $6.96, and GPT‑6 Sol’s max‑reasoning hits 61.4 at $1.26. Ultramarine’s max‑reasoning of 59.3 comes at $0.13, delivering near‑top performance at a fraction of the cost. Even the cheapest high‑performing model, GPT‑6 Luna, reaches 58.3 at $0.11, but its focus is on low‑reasoning tasks. Ultramarine’s blend of high scores and low price makes it the most cost‑effective option for professional knowledge work.
Frequently asked questions
What is the per‑task cost of Ultramarine 1 Flash compared to Claude Fable?
Ultramarine 1 Flash costs $0.13 per task, while Claude Fable 5 (max) costs $8.19 per task.
How does Ultramarine 1 Flash perform on the PI Knowledge Work benchmark?
It scores 59.3, three points below Claude Opus 5.5 (max) and ahead of GPT‑5.6 Sol (max) at $2.09 per task.
Does Ultramarine 1 Flash use fewer tokens than comparable models?
Yes, it generates 6.5 % fewer tokens than GLM‑5.3 Flash, dropping output from 32.7 M to 30.5 M tokens across six benchmarks.
The data positions Ultramarine 1 Flash as a disruptive force in the AI market for enterprise knowledge work. Its blend of high benchmark scores, token efficiency and a $0.13 per‑task price challenges the prevailing cost model of leading providers. As firms evaluate AI spend, Ultramarine offers a compelling alternative that could reshape budgeting for finance, legal and consulting automation. The next wave of adoption will test whether its cost advantage translates into real‑world productivity gains.
Sources: Hacker News, Paper Instruments Ultramarine page