Vellum - Flagship Model Report

2025 has been a defining moment for artificial intelligence. While breakthrough models, like the much anticipated release of GPT 5 , created huge waves in the AI space, leaders in the space are noticing clear redlining in performance capabilities with our current tech.

The US recently announced the Genesis Mission has formally kicked off a national effort to mobilize federal data, supercomputing resources, and national labs into a unified AI research platform. Its goal is to accelerate scientific and technological progress by making government datasets and compute directly usable by advanced models. In practice, Genesis marks the first major attempt to tie frontier AI capability to state-level scientific infrastructure and national priorities.

All the while leading AI researchers like Ilya Sutskever are amplifying this transition to research to see how AI progress can be achieved. In a recent interview , Ilya argued that the “age of scaling” is ending and that simply adding more compute won’t deliver the next order-of-magnitude breakthroughs. Instead, he describes a return to core research (e.g. new training methods, new architectures, and new ways for models to reason) as the real frontier from here.

Against this backdrop, the latest flagship model releases of GPT-5.1, Gemini 3 Pro , and Claude Opus 4.5  capture the tension of this moment: rapidly improving capabilities, rising expectations for national-scale impact, and a growing recognition that the next breakthroughs will come from deeper innovation. This report analyzes model performance across the board to see how each model provider is positioning itself, and what these shifts mean for the future of AI agents.

## Three trends you can’t ignore for 2026

Before diving into the numbers, it's important to contextualize the current landscape to understand where things are headed in 2026. These are the top three larger trends signaled by this new wave of flagship models.

### Shift to sophisticated, long-context agents

AI chatbots are yesterday’s story. These new models are signaling the rise of systems that can reason across massive context and execute complex, multi-step work. To see how dramatic this shift is, we need to look directly at the numbers driving it.

Since 2019, frontier model context windows have expanded by roughly three orders of magnitude, ~1,000 tokens to millions, leading some analysts to call this the “new Moore’s Law” of LLMs \[1\] \[2\] \[3\] . The moat right now being implementation, with around 62% of organizations still experimenting with AI agents. Out of these, almost two-thirds say they have not begun scaling AI across the enterprise, and fewer than 10% have scaled agents in any given function \[4\] \[5\] .

These massive improvements are pushing the  AI agents market to grow from with projections showing roughly $5.4 billion in 2024 to $7.6 billion in 2025, on track to reach about $47 billion by 2030 at a 45.8% CAGR \[6\] \[7\] \[4\] . AI budgets are

... (truncated, 22 KB total)