Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Tibo explains the 300k choice as a better tradeoff after evaluation. Running 1M consumes more usage.
- The reply lists no model, client, version, or plan scope. It cannot be treated as a new limit for every interface.
- The original discussion's 29% saving is a user's claim about Claude, with no evidence that a Codex change would save the same proportion.
- Keep the goal, relevant files, and necessary decisions, then remove unrelated input. No reset, extra allowance, or fixed saving percentage was announced.
The reply explains a default tradeoff
On October 10, Tibo @thsottiaux replied to a long-context discussion about why Codex uses 300k. He said the team had evaluated the choice and considered it better; 1M can also run, but consumes more usage. The point was an appropriate default, not a new 1M entitlement for all users.
The reply gives no applicable model, client, version, or plan list, and no evaluation method, scores, or usage multiplier. Keep 300k as the number in this explanation rather than substituting it for the context rules shown in every interface.
Why does longer context not necessarily offer better value?
Context is the input a model can refer to while working on a task, including the question, code, documents, and prior conversation. A larger context can hold more material, but fitting material into it is different from that material helping the current task.
Tibo explicitly says running 1M consumes more usage. He does not quantify the increase or say that tasks must become slower or results worse. Assess what problem the extra material solves rather than assuming a larger number improves usage efficiency.
If 1M can run, how do you know whether you can use it?
Check the model and client you actually use. Distinguish what the model can run, what that interface permits you to configure, and the default used by the current task. These can differ. A product leader saying 1M can run does not establish identical settings across desktop, CLI, cloud tasks, and the API.
Before changing configuration, check official documentation and the actual options for that interface. The reply gives no parameter name, procedure, or minimum version, so it does not establish a 1M setup recipe. Until you confirm applicability, keeping the current configuration makes results easier to assess than copying someone else's parameters.
Why is the claimed 29% saving not a Codex result?
The preceding post was from user rohit3a, discussing Claude and claiming that setting the autocompaction trigger to 400k could save 29% of weekly usage. That is a user's discussion of another product, not a Codex test result published by Tibo. The image and complete testing conditions were not verified in this review.
It therefore cannot become “Codex at 400k saves 29%,” and the preceding /autocompact idea cannot be treated as a universal Codex command. An autocompaction threshold concerns when existing input is condensed; context capability concerns the range an interface can process. They are different concepts.
What can you do without changing parameters?
Organize the material needed for the current task: the goal, relevant files, decisions that must be retained, and acceptance conditions. Remove unrelated logs, repeated output, and obsolete discussion from the current input. This is an input-organization suggestion, not a setting proven to save a specific percentage.
Here is a hypothetical example: you are editing login-page copy but also have a week's payment-troubleshooting logs. Keep the page files, wording requirements, and existing decisions; leave out unrelated logs for now. Add them when payment troubleshooting becomes the task. This controls the information scope without making expansion to 1M the first step.
When comparing usage, record the same model, similar tasks, and approximately comparable input. Different tasks at different times cannot be attributed accurately to context settings from a weekly balance alone. This clarification also announces no usage reset, extra allowance, or new Day number.
Sources
Frequently Asked Questions
Can 300k or 1M be converted directly into a Chinese-character count?
No. These figures should not be treated as character counts. Models generally measure input in tokens, and text, code, and different languages can be split differently. Check the applicable model and interface for the actual available range.
If I split the material into a new conversation, does it inherit earlier decisions?
The name of a new conversation does not establish that. When reorganizing a task, explicitly include the goals, key decisions, and files that must be retained, then check whether they are sufficient to continue. This article does not promise that any particular client automatically transfers all history.