OpenAI's advanced reasoning models are significantly underperforming on specialized tasks, with new benchmark results showing that domain-specific AI systems outpace general-purpose models by wide margins. When tested on smart contract security detection, OpenAI's o3 model achieved only 10.6% accuracy, while a specialized auditing system reached 87.7%, raising fundamental questions about whether the industry's investment in universal reasoning models is missing the mark .

Why Are General-Purpose Reasoning Models Struggling on Specialized Tasks?

The performance gap became visible when Cecuro, an AI-powered smart contract auditing platform, published results from EVMBench, an open-source security benchmark developed by OpenAI, Paradigm, and OtterSec . Cecuro's specialized multi-agent security system identified 101 of 120 high-severity vulnerabilities across 40 real-world audit cases, achieving an 87.7% detection rate. In stark contrast, OpenAI's o3 model detected just 10.6% of the same vulnerabilities.

The disparity extends across the entire AI industry. Anthropic's Claude Opus 4.6 scored 45.6%, while OpenAI's GPT-5.3-Codex and GPT-5.2 both achieved 39.2% on the detection task . Even Google's Gemini 3 Pro, at 20.8%, substantially outperformed o3. The pattern reveals a fundamental limitation: general-purpose reasoning models, despite their broad capabilities, lack the structured methodology and domain knowledge required for systematic vulnerability detection in blockchain systems.

The cost of this gap is substantial. Cecuro's earlier benchmark evaluated 90 real-world exploited contracts representing $228 million in losses and found a 92% detection rate, with Cecuro's system covering $96.8 million in exploitable value compared to $7.5 million for a standard frontier AI agent . This difference translates directly to financial protection and security outcomes.

What Architectural Differences Explain the Performance Gap?

The reason for the disparity lies in how these systems are built. Reasoning models like o3 bring strong logical capabilities but lack deep understanding of specific domains. For smart contract security, this means missing critical patterns that drive real-world losses . The specialized knowledge required includes lending protocol mechanics, automated market maker (AMM) price manipulation vectors, cross-contract callback risks, and DeFi-specific interaction patterns that general-purpose models simply haven't learned.

EVMBench itself uses containerized environments and deterministic testing to measure real-world performance, with all 120 findings being high-severity vulnerabilities independently confirmed through competitive audit processes . This rigorous evaluation methodology reveals that general-purpose models, when applied to specialized domains, fall far short of purpose-built systems.

How to Select AI Models for Specialized Work

• Domain-Specific Benchmarks: Evaluate models on tasks directly relevant to your industry rather than general knowledge tests. EVMBench, for example, measures real-world performance on blockchain security using actual audit cases rather than synthetic scenarios.
• Architectural Alignment: Assess whether a model's design matches your specific problem. General-purpose reasoning excels at broad tasks but struggles with specialized workflows requiring deep domain knowledge and structured methodologies unique to your field.
• Real-World Validation: Prioritize models tested on genuine cases from your industry. Cecuro's benchmark used 120 high-severity findings from 40 audit cases sourced from competitive platforms, providing authentic validation rather than theoretical performance metrics.

Meanwhile, OpenAI is responding to these limitations by building specialized models. The company announced GPT-Rosalind, designed specifically for life sciences research, marking the first release in its life sciences model series . Named after pioneering DNA researcher Rosalind Franklin, the model is now available as a research preview through OpenAI's access program to qualified customers including Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.

GPT-Rosalind addresses a concrete pain point in drug discovery. The typical timeline from target discovery to regulatory approval takes 15 years in the United States, with progress hampered by the difficulty of underlying science and fragmented research workflows . Scientists must navigate large volumes of literature, specialized databases, experimental data, and evolving hypotheses to generate and evaluate new ideas.

"This is the first release in our life sciences model series and we view it as the beginning of a long-term commitment to building AI that can accelerate scientific discovery in areas that matter deeply to society, from human health to broader biological research," OpenAI stated.

OpenAI, statement on GPT-Rosalind

This strategic shift reflects broader changes at OpenAI. The company is increasingly focusing on business-oriented products and specialized models rather than pursuing a single universal reasoning system . Sarah Friar, OpenAI's chief financial officer, revealed that the company is developing a new model codenamed Spud, described as its "smartest model yet" with "stronger reasoning, better understanding of intent and dependencies, better follow-through and more reliable output in production" .

Financial pressures are driving this pivot. Business customers now account for 40% of OpenAI's revenue, up from 20% when Friar joined in 2024, and the company expects this to reach 50% by year's end . The shift away from consumer products like Sora, the AI video generator, underscores the priority. Friar acknowledged the difficult choice: "I think it was a little heartbreaking, but we're like, OK, it's not the main event right now. We need to make sure that our new model that's coming has enough compute."

Friar

The broader implication is that the AI industry may be entering a new phase. Recent benchmark results and product announcements suggest that specialized systems may deliver superior performance on domain-specific tasks, prompting companies like OpenAI to invest in specialized models alongside general-purpose systems. OpenAI's o-series reasoning models remain powerful tools, but their limitations on specialized tasks suggest that the future of AI may belong to systems designed with specific problems in mind, not universal solvers attempting to excel everywhere at once.