Ultramarine 1 Flash

Developed by Paper Instruments, Ultramarine 1 Flash is an open-weights model engineered specifically for high-stakes professional environments. It delivers Fable-class performance across critical knowledge-work domains, including operations, consulting, law, and finance.

"Ultramarine 1 Flash is designed for professional work, bridging the gap between efficiency and elite-level expertise."

🚀 Availability & Resources

• Ultramarine API $\rightarrow$ Available Now
• Technical Report $\rightarrow$ Read More
• Model Weights $\rightarrow$ Available Now Coming Soon

---

📊 Performance Benchmarks

Ultramarine 1 Flash competes at the highest level of professional AI. Below is the performance breakdown across six key industry benchmarks:

---

⚡ Token Efficiency

Ultramarine 1 Flash doesn't just work better; it works faster by being more concise. Compared to its base model, GLM-5.3 Flash, it achieves superior results while generating fewer tokens.

Token Consumption Comparison:

• GLM-5.3 Flash: $32.7\text{M}$ tokens
• Ultramarine 1 Flash: $30.5\text{M}$ tokens
• Net Reduction: $\Delta = -6.5\%$

(Note: Where GLM-5.3 Flash scores are public, token counts are derived from internal runs approximating those scores.)

---

💰 The Cost-Quality Frontier

Ultramarine 1 Flash redefines the value proposition for professional AI, offering elite performance for a fraction of the cost.

Economic Comparison (PI Knowledge Work)

At a cost of $0.13 per task, Ultramarine 1 Flash outperforms:

1. Claude Fable 5 (max): $8.19/task
2. GPT-5.6 Sol (max): $2.09/task

It remains highly competitive with top-tier models, trailing Claude Opus 5.5 ($6.96/task) and **GPT-6 Sol** ($1.26/task) by fewer than three points.

API Pricing Structure

---

🛠️ Deep Dive: PI Knowledge Work

The PI Knowledge Work benchmark is a proprietary internal evaluation focusing on real-world professional outputs.

What it tests:

• Generation of finished Excel workbooks
• Creation of PowerPoint decks
• Drafting of Word documents
• Integration with live business systems (e.g., ERP and MES)

Evaluation Methodology: Each task is graded via a strict rubric based on:

• Substance
• Accuracy
• Presentation

The validity of this benchmark is supported by its strong correlation with the Artificial Analysis Intelligence Index: $r = 0.81$

---

🖋️ Client-Ready Output

Ultramarine 1 Flash is not just a reasoning engine; it is crafted with a "sense of client readiness." This means the model avoids fluff and focuses on producing:

• Clean prose
• Simple structures
• Straightforward professional communication

{
  "model": "ultramarine-1-flash",
  "prompt": "Analyze the Q3 fiscal report for discrepancies.",
  "temperature": 0.2,
  "style": "professional_concise"
}