OpenAI has announced improvements to Codex and ChatGPT Work that are intended to increase the effective usage the user receives. According to the announcement, the company reuses context between subagents, keeps context cached for longer and improves request routing to increase cache hits. The announcement was published on 17 October in Israeli time and specifies neither a uniform saving nor a new quota for each plan.
The implication described is a better use of information that the system has already processed, rather than starting every time from the same point. There is no guarantee that each request will use the cache or that each task will consume less resources. The benefit may depend on the nature of the job and the available context. It is therefore appropriate to distinguish between an improvement in the service infrastructure and an explicit change in the subscription terms or price.
What do the three changes include?
Reuse in context between sub-agents refers to information that agents can rely on during their work. Longer caching keeps material that has already been processed available for reuse for an additional period. Improved routing seeks to route requests in a way that increases the possibility of using this material. These are related components, but each deals with a different part of the process that the company describes.
The message does not say that the cache is a permanent memory that the user can edit, or that information from each account is shared by each agent. It also does not describe a change in access permissions to documents or tools. Therefore, there is no reason to interpret reuse in context as extending access to data not previously granted. Permissions, product memory, and infrastructure cache are separate issues to be checked according to the relevant configurations.
What should be measured in actual use?
A team using Codex can see how much valid work was completed in a similar period of use, along with the number of attempts and waiting time. A comparison between very different tasks can be misleading even if the expenditure appears to be lower. It is advisable to keep the nature of the task and the conditions of success, to check if the improvement is noticeable in the work process itself and not only in a partial index shown along the way.
The announcement does not instruct users to change code, switch plans or remove controls to receive the improvement. Anyone who manages a budget should keep checking the account data and the rules that apply to it. Improved resource use does not promise unlimited access and is no reason to ignore quota or billing notices. The practical test is the ratio between the resources used and the results completed with the required quality.
Even after the improvement, the division of labour between agents must be appropriate to the mission. If one result depends on another, reuse in the context does not cancel the dependency. It's best to pre-define responsibilities and expected output, and check the combined result. This allows teams to benefit from infrastructure efficiency without assuming that every task suits delegation or that every output is ready for use.