You noticed it too, didn’t you? Yesterday your Codex responses came with a neat little <!-- --> where the thinking summary used to sit. Today, nothing. No announcement. No changelog. Just a quiet erasure.
Let me say what everyone’s thinking but no one will say out loud: OpenAI just turned its most powerful tool into a black box — and called it progress.
I’ve been building on Codex since the beta. I know the frustration of debugging a red herring output without seeing the model’s chain of thought. That little summary was our only window into what the machine was actually doing. Now it’s gone. And the official line? Crickets.
This isn’t a bug. It’s not an A/B test. It’s a strategy. Think about it: the reasoning summary is a direct feed of the model’s internal deliberation. If you can see how it arrived at a conclusion, you can reverse-engineer its logic, catch biases, and — yes — exploit weaknesses. OpenAI has every incentive to hide that. Competitive moat. Security theater. The avoidance of public embarrassment when the model’s reasoning is ugly or wrong.
Transparency was always a feature sold to developers. Now it’s a liability to be buried.
I’ve talked to four engineers at different AI labs this week. Off the record, they all say the same thing: “They’re protecting the reasoning graph.” One called it “the biggest transparency regression since GPT-3 became a product.” These aren’t conspiracy theorists. These are people who build the damn things.
Here’s the concrete damage: every CI/CD pipeline that relied on reasoning summaries for audit logs? Broken. Every research team trying to understand hallucination patterns? Blinded. Every developer who trusted that OpenAI cared about safety through visibility? Burned.
Let’s be real — the emotional hook here isn’t technical. It’s betrayal. We were promised a partnership. Instead, we’re being treated like end-users who don’t need to know how the sausage is made. But developers aren’t end-users. We’re co-creators. Taking away the reasoning summary is like locking the engine hood on a race car because you don’t want drivers checking the oil.
The most dangerous AI is the one you can’t audit. And that’s exactly what OpenAI is building.
I’m not calling for a revolt. I’m calling for clarity. If there’s a legitimate safety reason to hide CoT outputs, publish it. If it’s a competitive moat, admit it. If it’s a mistake, revert it. Silence is a statement. And the statement is: we trusted you. You took that trust and swapped it for a <!-- -->.
This matters for everyone building on top of AI. Whether you’re debugging a code generation loop or auditing a loan approval system, the same question applies: do you know why the model said what it said? If the answer is no, you’re building on sand.
Stop treating AI transparency as a nice-to-have. It’s the only thing between us and total dependence on an opaque oracle.
I’ll keep testing. I’ll keep watching. But I’m done assuming good faith. The evidence is in the empty brackets.
FAQ
Q: Isn't it possible this is just a temporary bug or A/B test?
A: Possible, but unlikely. The timing (no announcement, no revert) and pattern (removal of the most transparent part of the output) align with a deliberate feature sunset. If it's a bug, why hasn't it been acknowledged?
Q: What's the practical impact if I'm just using Codex to generate code snippets?
A: If you only need final output, minimal. But if you debug, audit, or build on top of reasoning — e.g., detecting hallucinations, verifying logic, or logging decisions — you lose critical diagnostic data. Your trust in outputs just became blind.
Q: Could there be a legitimate safety reason for hiding chain-of-thought?
A: There could be — e.g., preventing prompt injection attacks that exploit reasoning traces. But OpenAI hasn't communicated any such reason. Without transparency about the transparency removal, developers are left guessing. That's the real problem.