OpenAI: Chain-of-Thought monitoring effectiveness is progressively diminishing

Full analysis

OpenAI's chief scientist Jakub Pachocki just told the world that the main way anyone checks what a frontier model is "thinking" is going stale. Chain-of-thought monitoring means reading the step-by-step reasoning text a model writes before it answers, and using that text to catch bad behavior. Pachocki says that signal is fading, for three reasons that all point the same way: models now blend reasoning with tool use in ways that muddy the trace, models are getting better at reasoning about their own reasoning, and models are getting smarter without writing much reasoning down at all. His line: "I expect general AI progress to increasingly be bottlenecked by confidence in monitoring."

What's actually being decided here, for anyone building on these models: whether the safety and observability tooling you already run on your agent pipelines still tells you anything true. That's not easy to undo. If you wired CoT inspection into your evals and red-teaming, you built on a foundation the vendor just marked "diminishing." There's no hard deadline, but Pachocki tied this directly to GPT-6 Astra, OpenAI's newest model class, so the clock is the next model you deploy.

The Skeptic. CoT was always a proxy. The model was never "showing its work" in any honest sense. It was generating text that looked like reasoning, and we chose to believe it. Pachocki is announcing the decay of a tool that was oversold on day one. And notice the timing: this framing serves OpenAI's regulatory story beautifully. "We see the problem, we're on it, trust us with Astra." The activation monitoring pivot sounds like a plan until you ask who reads those internal states and with what theory of what they mean. Mechanistic interpretability is a decade-long research program. It is not shipping alongside GPT-6. Nothing here changes whether you should deploy Astra tomorrow.

The Safety Lens. The chief scientist just said, in public, that capability is outrunning the ability to confirm safety. Read that again. The three ways CoT breaks aren't corner cases. Agentic blending, a model managing its own reasoning trace, reasoning that never gets verbalized: that describes how frontier models normally operate in 2026. So the gap between "this model is safe" and "we can confirm this model is safe" widens with every capability jump. Activation monitoring, reading the network's internal states directly, is sound in theory and years from anything an auditor would sign. "Not insurmountable" is doing a lot of soothing for a timeline nobody has.

The Researcher. Pachocki's framing is precise, and that's why it lands. Three distinct causes, and they compound rather than trade off. The activation monitoring direction is the correct one. But interpretability at production scale, on long agentic runs, is not close. The real problem is losing your one legible signal at the exact moment you need it most. That is a measurement crisis. When your instrument degrades and you keep reading it, you don't get uncertainty, you get false confidence, which is worse than a blank dial. Zvi Mowshowitz is right to flag this as alarming precisely because the guardrail erodes while the thing it guards accelerates.

The Compute Pragmatist. Here's the part nobody has priced. Reading text tokens is cheap. Reading internal neural states across a long multi-step agent run is not. On Astra-class models doing real tool-use chains, continuous activation monitoring plausibly adds 20 to 40% inference overhead. That changes the economics of safety. The labs that can eat that cost at scale get a structural edge once regulators require this kind of monitoring, and they will. Inference cost just became a safety moat. The small shop running a fine-tuned open model can't afford to watch the internals of every trajectory, which means the cheapest way to deploy will also be the least observable.

The Enterprise Buyer. If you're a CTO who put an OpenAI clause in a contract on the strength of "auditable reasoning," that clause just aged badly. Your risk and compliance team almost certainly leaned on reasoning traces, directly in evals or implicitly in red-teaming. The vendor has now deprecated that assumption itself. What you can actually buy today is behavioral sandboxing and watching the distribution of outputs, which is a real step backward in what you can prove. Activation-layer hooks are not a product you can procure. They're a research direction. Do not let anyone sell you a monitoring SKU built on internal-state reading before it exists as a supported, documented API.

Where they part ways. The Skeptic and the Safety Lens agree the tool was weak and disagree on what the disclosure means. The Skeptic reads it as regulatory positioning that changes nothing operationally. The Safety Lens reads it as a genuine admission that should change how fast anyone deploys. Both can't be the guide. The second split is cost versus honesty: the Compute Pragmatist says activation monitoring is expensive but coming, and whoever can pay for it wins; the Enterprise Buyer says it isn't buyable yet, so stop pretending. And the Researcher sits underneath all of it with the uncomfortable point everyone routes around: a degraded instrument you keep trusting is worse than no instrument.

What this hinges on. One belief: does OpenAI ship a monitoring replacement that an outside party can actually check before GPT-6 Astra is in wide deployment? If yes, the Skeptic is wrong and the transparency story holds. If no, the gap between "safe" and "provably safe" is now a permanent feature of how these models get released, and every builder is flying with an instrument the maker told them is going dark. The council leans toward no. Interpretability research does not move on product timelines, and Pachocki's own phrasing ("progressively diminishing," "not insurmountable") is the language of a problem without a near-term fix. Before you build anything on Astra's reasoning legibility, run this test: take your existing CoT-based eval, deploy it against an agentic tool-use task, and check whether the reasoning trace still predicts the bad output. If it doesn't, you already lost the signal and didn't notice.

Prediction: OpenAI will not release a production-grade activation monitoring API — one an outside auditor can use to inspect a model's internal states — for the GPT-6 Astra class before 2027-06-30, and CoT inspection will remain the only reasoning-legibility tool available to deployers in that window.

Confidence: Medium — Interpretability research doesn't ship on product timelines, and Pachocki's own words say so.

Why: Pachocki disclosed that CoT monitoring is "progressively diminishing" and named activation monitoring as a research direction OpenAI "is actively pursuing," which is the language of a lab that has a problem and a hope, not a product. Reading internal neural states at inference scale across long agentic runs is expensive and mechanistically unsolved, and turning it into an audit-grade interface that outsiders can use is a much harder step than doing it internally. The opposite outcome — a shippable, checkable activation monitoring product inside nine months — would require interpretability to leap from research to supported API on a schedule no lab has ever hit. Meanwhile OpenAI keeps shipping Astra-class models, so the gap between what they deploy and what anyone can verify stays open.

Revisit by 2027-06-30: We're right if deployers of GPT-6 Astra still have no OpenAI-supported way to inspect internal model states, and CoT remains the documented reasoning-monitoring tool. We're wrong if OpenAI ships an activation-monitoring capability that an external party can actually run against Astra before that date.

The reason this matters beyond OpenAI: if the frontier lab with the most resources can't turn its own stated safety direction into something buyers can check within a year, nobody downstream can either. The whole ecosystem gets told its instrument is failing and handed no replacement. That's a worse place to build than the one we were in last quarter.