OpenAI Codex Head Tibo Publicly States Team Chooses 300k as Codex Default Context Length
Codex Head Tibo: 300k context is a better default value after data validation
OpenAI Codex Head Tibo publicly stated that the team chose 300k as the default context length for Codex because data analysis showed that this value performed better.
He pointed out that the system can also run with a 1M context, but it will consume more usage quota, so the default value needs to balance performance and cost.
Previously, Codex had limited the product layer context of GPT-5.6 Sol to around 272k, then briefly adjusted it to 372k before retracting, as the higher window led to subscription user quota consumption exceeding expectations.
Tibo had previously disclosed how to increase the context of models supporting a 1M window (like GPT-5.6 Sol) to 1 million through configuration files, setting an automatic compression threshold of about 900k, but emphasized that the default value has been fine-tuned.
A larger context can retain more code, tool outputs, and conversation history, reducing premature summaries, but will significantly increase the quota consumption per request, especially after exceeding about 272k, triggering higher billing rates.
Users can actively choose a larger window through configuration, while the default 300k is aimed at controlling costs for most scenarios; subscription user quotas are inclined towards efficient use of the default, while users who actively open a larger window consume faster, and the platform manages overall usage efficiency through the default value.
OpenAI model documentation shows that some cutting-edge models natively support a window of about 1.05 million tokens, with differences between product layer default values and API limits.
Source: Public Information
ABAB AI Insight
Tibo has previously publicly adjusted the Codex context limit and explained the abnormal quota consumption, having raised the window from 272k to 372k before retracting due to rapid consumption, while also pushing the 1M option open to ChatGPT accounts. His approach is to control costs through product layer restrictions rather than the model's capabilities.
OpenAI has invested engineering and product resources into optimizing the default value, allowing most users to achieve a balance of cost and effect without actively configuring, while retaining configuration interfaces for heavy users to choose larger windows, thus separating high consumption behavior from the default path.
This is similar to Claude's product layer default limits on long contexts and the gradual opening path for context length in early ChatGPT; the current programming agent is transitioning from "the bigger, the better" to "the default is optimal," where quota and latency have become more practical constraints than the native window.
Essentially, this is a redistribution of pricing power and usage efficiency, with the platform controlling most quota consumption within an efficient range through the default value, while users needing longer contexts must actively bear higher consumption, making marginal costs explicit.
ABAB News · Cognitive Law
The default value is the intersection of cost and effect, not the capability limit
Being able to open to a larger size does not mean the default should be larger
Quota consumption becomes a constraint earlier than window numbers