Andrej Karpathy, a former OpenAI co-founder, warned the AI industry that prioritizing agent development over mastering foundation models is a mistake, citing OpenAI’s failed attempt to automate tasks in 2016 which delayed progress by five years. During a July 5 internal session, he emphasized that foundation models are the true core of AI progress. Karpathy asserted that independent developers currently rival tech giants in agent capabilities because major labs lack a multi-year head start in this specific field. He advised developers to focus on the underlying architecture rather than forcing current models into unreliable agentic roles, noting that bridging the gap between a demo and a functional product often requires a decade of development.

The industry’s struggle with foundation model limitations was highlighted on June 30 with the release of Anthropic’s Claude Science workbench and OpenAI’s GeneBench-Pro benchmark. GeneBench-Pro, which evaluates AI performance across 129 real-world scientific workflows, reported that the leading model, GPT-5.6 Sol, achieved only a 28.7% end-to-end pass rate under Max inference settings. The highest performing non-GPT model, Claude Opus 4.8, reached a 16.0% pass rate. These results expose a "notice-act gap," where models can detect anomalies but struggle to initiate correct analytical adjustments or downstream methodological changes.

In response to these structural challenges, the three major labs have adopted distinct ecosystem strategies. Anthropic’s Claude Science workbench integrates external vertical tools via the MCP protocol to manage complex scientific research processes. OpenAI is utilizing GeneBench-Pro to establish evaluation standards while promoting its specialized biological reasoning model, GPT-Rosalind, to enterprise partners. Google DeepMind maintains a competitive advantage by bundling its proprietary scientific models, such as AlphaFold and AlphaGenome, directly into its underlying infrastructure. Despite these efforts, pharmaceutical giant Novo Nordisk continues to trial multiple vendors' solutions in parallel, indicating that no single provider has yet secured a dominant position in the open scientific research market.

Andrej Karpathy, a former OpenAI co-founder, warned the AI industry that prioritizing agent development over mastering foundation models is a mistake, citing OpenAI’s failed attempt to automate tasks in 2016 which delayed progress by five years. During a July 5 internal session, he emphasized that foundation models are the true core of AI progress. Karpathy asserted that independent developers currently rival tech giants in agent capabilities because major labs lack a multi-year head start in this specific field. He advised developers to focus on the underlying architecture rather than forcing current models into unreliable agentic roles, noting that bridging the gap between a demo and a functional product often requires a decade of development.

The industry’s struggle with foundation model limitations was highlighted on June 30 with the release of Anthropic’s Claude Science workbench and OpenAI’s GeneBench-Pro benchmark. GeneBench-Pro, which evaluates AI performance across 129 real-world scientific workflows, reported that the leading model, GPT-5.6 Sol, achieved only a 28.7% end-to-end pass rate under Max inference settings. The highest performing non-GPT model, Claude Opus 4.8, reached a 16.0% pass rate. These results expose a "notice-act gap," where models can detect anomalies but struggle to initiate correct analytical adjustments or downstream methodological changes.

In response to these structural challenges, the three major labs have adopted distinct ecosystem strategies. Anthropic’s Claude Science workbench integrates external vertical tools via the MCP protocol to manage complex scientific research processes. OpenAI is utilizing GeneBench-Pro to establish evaluation standards while promoting its specialized biological reasoning model, GPT-Rosalind, to enterprise partners. Google DeepMind maintains a competitive advantage by bundling its proprietary scientific models, such as AlphaFold and AlphaGenome, directly into its underlying infrastructure. Despite these efforts, pharmaceutical giant Novo Nordisk continues to trial multiple vendors' solutions in parallel, indicating that no single provider has yet secured a dominant position in the open scientific research market.