Beating OpenAI at Adaptive Compute

The Challenge OpenAI Hasn't Solved

In a recent interview with Latent Space (August 15, 2025), OpenAI’s President Greg Brockman reflected on the challenge of making AI systems compute more intelligently rather than more expensively:

"Model switchers are not necessarily the future… it feels preferable to have a fully integrated system that just does the right thing. But evidence points toward a menagerie of models with different strengths and weaknesses… it's much easier to have a small, fast model that can generate a lot of tokens, coupled with a more expensive reasoning model. We haven't really cracked adaptive compute within the architecture yet—but doing it within orchestration is straightforward…"

That quote captures a key reality: even OpenAI, after billions of dollars of research, has not yet solved adaptive compute — the ability for AI to automatically decide how much thinking power to use depending on the problem.

Instead, their latest product, GPT-5 Router, is a manual workaround — a system that switches between models rather than one that adapts naturally. It's powerful, but not efficient.

At Concavity.ai, we took a different path — and built what OpenAI admits they couldn't.

Our Breakthrough

At Concavity.ai, we've achieved what OpenAI could not: native adaptive compute through a new agentic architecture that dynamically allocates computation across specialized models in real time.

Our system integrates a fast, lightweight model — built on Qwen3-30B-A3B-Instruct-2507 — with larger, more capable reasoning models such as DeepSeek V3.1 and DeepSeek R1. The small model handles the majority of tokens for speed and efficiency, while complex reasoning tasks are escalated automatically to the larger models. This architecture allows our system to think deeply only when needed — and stay fast when it's not.

We've also incorporated a multimodal vision-language model, enabling unified reasoning across text and image. On the GAIA validation benchmark, our system achieves 68.45% on the full (text + image) set — outperforming OpenAI's Deep Research Agent, which scores 67.36% on text-only tasks.

Even more striking: our system costs only hundreds of dollars to train, and our inference cost is roughly 100× lower than OpenAI's Deep Research Agent. With adaptive compute, we deliver performance that rivals the best — at a fraction of the cost.

Product: Deep Research Agent for College Students

We're applying this architecture to build the first affordable Deep Research Agent for college students.

Academic reasoning tasks are often too complex for single-model systems like GPT-5 or Gemini 2.5 Pro, yet agents like OpenAI's Deep Research or Google's Deep Think remain prohibitively expensive. Our adaptive-compute agent offers the best of both worlds — top-tier reasoning at student-friendly prices.

Our upcoming product will help students:

• Conduct deep academic research with multimodal context (text, figures, equations)
• Generate structured, citation-ready outputs
• Learn interactively through reasoning traces and self-explanations

With this technology, we're making world-class AI research assistants accessible to everyone.

The Future We're Building

The first era of AI proved that large language models could generate. The next era will prove that intelligent agents can reason, act, and learn collaboratively.

At Concavity.ai, we're building that future — a foundation for scalable, adaptive, and truly intelligent systems.

Stay tuned as we share more about our research, benchmarks, and early products in the coming weeks.