Foundation Models Brief: Agent Capabilities Expand, Efficiency Improves

What Actually Changed

Google is transforming Gemini into an AI agent capable of planning, executing tasks, and working across business applications and systems. The agent can delegate work to subagents, utilize multiple AI models, and has its own workplace identity with an email address. Alibaba released Qwen-Image-2.1-Turbo, an 8-step 7B image model that generates and edits images in 5x fewer steps than its predecessor. Anthropic announced Claude Corps, a national fellowship for extending AI benefits to communities across America, and revealed Claude Science helped produce the first complete map of the sky in UV light.

Why It Matters

The emergence of multi-agent systems represents a fundamental shift toward more autonomous AI systems that can coordinate complex workflows across business applications. This development addresses the growing need for AI systems that can handle long-horizon tasks requiring planning, execution, and revision. Meanwhile, efficiency improvements like Qwen-Image-2.1-Turbo's reduced denoising steps and Asana's 76x cost reduction with GPT-6.1 Sol demonstrate that advanced capabilities are becoming more accessible and cost-effective, potentially accelerating enterprise adoption.

Capability and Benchmarks

Research continues to advance model reasoning capabilities, with new approaches like LoopOPD for cross-loop on-policy distillation in Looped Language Models showing promise for parameter-efficient scaling. The Saluki 27B model demonstrates that quantization can enhance specific capabilities like tool calling while maintaining competitive performance. However, studies reveal limitations in using LLMs for machine translation quality annotation, showing both humans and models face similar challenges in evaluation tasks. These developments highlight the ongoing need for specialized approaches to different reasoning and evaluation tasks.

Compute, Cost and Pricing

Enterprise implementations are showing dramatic cost efficiency improvements, with Asana achieving 76x cost reduction and 5x speed improvements in browser tests using GPT-6.1 Sol. Meanwhile, Alibaba's Qwen-Image-2.1-Turbo offers developers a more efficient image generation model with 5x fewer denoising steps. These improvements suggest that the industry is making significant progress in reducing computational requirements while maintaining or enhancing capabilities, potentially lowering barriers to entry for advanced AI applications.

What We Are Watching

We're closely monitoring the development of normative competence in LLMs, particularly how models can discern community norms without relying on static pretrained knowledge. The introduction of stochastic teacher intervention for agentic on-policy distillation presents another promising approach for improving multi-turn agentic tasks. Additionally, the Plan-and-Patch framework for diffusion language models offers an innovative approach to agentic planning that repairs rather than regenerates entire plans, which could significantly improve efficiency in complex task execution.