OpenAI opens GPT-5's raw reasoning to external testers

OpenAI is letting a select group of independent labs peek under the hood of its unreleased GPT-5 model, granting them direct access to the system’s chain-of-thought reasoning. This marks a significant shift in transparency, designed to catch deceptive behaviors like scheming or sandbagging that might be invisible if you only look at a model’s final output. The move is part of a broader external testing regime the company details in a new blog post, which breaks down how it uses third-party assessors to validate safety claims before shipping frontier AI.

The company says its external collaborations take three forms: independent evaluations of risky capabilities like biosecurity and cybersecurity, reviews of the methodologies it uses to assess risk, and direct probing by subject-matter experts. For GPT-5, OpenAI coordinated assessments with organizations like METR and SecureBio, which ran the model through tests measuring long-horizon autonomy and virology planning feasibility. Crucially, testers got access to early model checkpoints with fewer safety mitigations, allowing them to probe the underlying capabilities without guardrails getting in the way.

Providing external labs with chain-of-thought access is a big deal. It’s essentially letting someone read the model’s internal monologue as it works through a problem. The stated goal is to identify sandbagging—where a model hides its true abilities—or scheming behavior that might only be discernible in those reasoning traces. It’s a level of access that carries significant security implications, and OpenAI acknowledges it applied security controls that it plans to evolve as the tech improves.

This isn’t charity. By publishing summaries of these pre-deployment evaluations in system cards and supporting assessors in releasing detailed findings, OpenAI is trying to build trust through transparency. The bet is that showing your work—warts, reasoning traces, and all—is better than asking the public to trust a black box. Whether this becomes the standard for the industry or a one-off flex remains an open question, but for now, it’s a concrete example of what third-party AI auditing actually looks like in practice.