# Microsoft AI — Optimizing the frontier performance curve

- Company: Microsoft AI (microsoft.ai)
- Announced: 2026-07-29
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://microsoft.ai/news/optimizing-the-frontier-performance-curve
- Record: https://forck.live/items/8418-optimizing-the-frontier-performance-curve
- Title in English (translated by forck.live from en): Optimizing the frontier performance curve
- Subject: MAI
- Models affected: MAI-Cyber-1-Flash, MAI-Code-1-Flash, MAI-Image-2.5-Flash, MAI-Voice-2-Flash, MAI-Transcribe-1.5, MAI-Image-2.5-Pro, GPT 5.4, GPT-5.4 Mini, Claude Haiku 4.5, GPT-5.6, GPT-Image-2

Microsoft AI released MAI-Cyber-1-Flash, optimized for the MDASH harness, achieving 96% on CyberGym at 50% of the cost of the previous best MDASH offering, and served on H100s. The post also announces MAI-Code-1-Flash, MAI-Image-2.5-Flash, MAI-Voice-2-Flash, and MAI-Transcribe-1.5, each deployed in Microsoft products with reduced GPU costs and improved performance. These models are part of a broader effort to optimize frontier performance against cost by co-optimizing models, harnesses, and reinforcement learning environments.

## Evidence

Verbatim from https://microsoft.ai/news/optimizing-the-frontier-performance-curve:

> This week we released MAI-Cyber-1-Flash optimized for our MDASH harness. Together, the system delivers 96% on CyberGym (on the benchmark's any crash score; outperforms Mythos on the CyberGym leaderboard) at 50% of the cost when compared against our best offering in MDASH today.

---

Record: https://forck.live/items/8418-optimizing-the-frontier-performance-curve
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md