OpenAI cuts chain-of-thought extraction to 0.2%, closing the model-copying shortcut
OpenAI's new defences cut reasoning-extraction success from 54.0% to 0.2%, forcing firms to rethink AI costs built on reselling frontier-model output.

OpenAI has moved to lock down the "thinking" behind its models. In internal black-box red-team tests, the success rate of extracting reasoning content fell from 54.0% to 0.2% under its updated safeguards, and from 22.9% to 0.9% for current reasoning requests. For companies whose AI strategy rests on distilling a cheaper model from a frontier one, the arithmetic has changed.
Why the thinking process is now behind a lock
The answer is simple: a chain of thought is itself an asset that can be distilled.
OpenAI's notice defines adversarial distillation as using a frontier model's output, without authorisation, to train another model so that it gains similar capabilities without the corresponding safety protections. The concern is not copying as such, but that the safety constraints are lost along the way. If high-risk capabilities in areas such as cybersecurity and biology are transferred to a less restricted model, the risk doubles.
The defences come in several layers: adversarial training so the model recognises and refuses requests to extract reasoning chains; input and output classifiers to detect attacks and leaks; and API limits on reusing reasoning content across sessions, models and organisations. GPT-6.2 Astra and GPT-6.2 Sol also bind protected reasoning content to the conversation history that produced it, discarding later reasoning if earlier messages are altered. Older models cannot read the protected reasoning content of these two new models.
There is enforcement too. OpenAI says it has taken action every week this year against accounts involved in unauthorised distillation, at a scale ranging from thousands to tens of thousands, with a marked increase since the summer.
When a leading vendor starts treating its thinking process as an asset to be locked, the business of living off resold model output reaches its end.
The counter-argument: does closing up bind smaller firms tighter?
This is the most common objection, and the logic sounds smooth: the tighter capabilities are held, the more bargaining power concentrates upstream, and small companies can only accept the price.
The worry is reasonable, but it rests on a premise: that using someone else's model is your core capability. If a company's AI competitiveness is built entirely on calling an API and repackaging the result, it had no moat to begin with. Whether the door is locked is only a matter of time.
Look at it the other way: what OpenAI has locked is the reasoning chain, not ordinary calls. Companies using model capabilities through the API for their own business systems are not the target. What is blocked is the attempt to reverse-engineer a cheap substitute from a small amount of high-quality output.
When does this judgment fail? If a company's business depends heavily on reprocessing a specific model's output and it has no data, scenarios or process experience of its own, it will indeed suffer. That is not a wrong judgment; it is a dangerous position to stand in.
What changes for business owners
In operational terms, what really changes is how AI spending is calculated.
Many owners used to compute "model unit price × call volume", assuming that finding a cheaper substitute model would keep pushing costs down. Now another item must be added: the legality and stability of the source of capability. Once the upstream tightens, capabilities obtained through distillation can be cut off at any time, and earlier investment is lost.
Seen another way, this pulls competition back to the scenario itself. With the same model capability, whoever can plug it into their own business processes and capture real customer enquiries and orders is the one whose value holds. Shu Yi, executive director of Sheyu Jiye, puts it in a way we agree with: AI will not make decisions for the boss; it puts the numbers the boss should see in front of them first. However strong the model, it ultimately comes down to how the business runs.
One thing to do now
List the AI tools you are currently using and ask of each one: whose capability does it depend on, and if that capability were tightened tomorrow, would my side stop working?
If the answer is yes, it is time to move a step toward your own scenarios and data. For companies in foreign trade and independent sites, for example, rather than spending energy hunting for cheap models, it is better to first make the official website a solid entry point that can be found by search and can capture enquiries. What a tool like SHEYU AISITE can do is build a site by drag-and-drop without coding, start from templates, and handle basic SSR and SEO configuration so that content can be crawled and enquiries land in the back office for follow-up. What it cannot do is guarantee that you are indexed or ranked, or replace the competitiveness of the product itself. Tools solve the entry problem; the business still depends on people.
The stronger the model, the more you should spend where others cannot copy you: your scenarios, your customers, and the processes you run every day.