A significant reduction in the context window for several models, including the GPT-5.x series, has been discovered in the openai/codex repository, which may indicate a shift in OpenAI's strategy toward optimizing inference costs.
What Happened
In the release/0.144 branch of the openai/codex repository, via pull request #33972 merged on July 18, 2026, model metadata was updated. This resulted in a reduction of the available context window from 372k to 272k tokens, representing a decrease in memory capacity of approximately 27%.
Context
The changes affected OpenAI's infrastructure repository and were recorded in the metadata for GPT-5.x models. Such parameter adjustments during the development of new model generations are often linked to the need to reduce computational complexity when working with long sequences.
Why It Matters for the Industry
For the industry, this could signal a transition from a strategy of providing maximum context lengths to an "efficient context" model. This may stimulate the development of prompt compression tools, more advanced RAG (Retrieval-Augmented Generation) methods, and specialized summarization layers to compensate for the reduction in the model's direct memory capacity.
Why It Matters for Users
Developers using APIs or tools based on Codex and GPT-5.x need to conduct an immediate audit of their current systems. There is a risk of "context length exceeded" errors occurring in existing pipelines. A review of long-prompt management strategies and optimization of data processing architectures will be required to operate within the new 272k token limit.
What Remains Unknown / Limitations
The exact reasons for the parameter change (whether architectural optimization or economic feasibility) have not been officially confirmed.
Sources
Author
Look at AI, Editorial Staff