OpenAI has quietly reduced the working context window for its Codex coding agent and the underlying GPT-5.6 model family, cutting the usable limit from roughly 372,000 tokens to 272,000 tokens. The change, surfaced through a merged pull request on the company's open-source Codex repository, has drawn sharp reactions from developers who rely on large contexts to analyze entire codebases in a single pass.
The adjustment reflects a familiar tension in the AI industry, where companies balance the cost of long-context inference against the convenience it offers users. For the latest AI industry coverage, the move underscores how even small changes to model metadata can ripple across millions of developer workflows.
What changed and where it was spotted
The reduction became visible in the Codex repository's bundled model metadata file, `codex-rs/models-manager/models.json`. A pull request titled "Backport refreshed bundled model metadata to 0.144," authored by OpenAI engineer sayan-oai and merged on July 18, 2026, refreshed the metadata that stable Codex clients use to understand each model's capabilities.
The current file shows a `context_window` value of 272,000 tokens across the GPT-5.6 lineup, including the GPT-5.6-Sol, GPT-5.6-Terra, and GPT-5.6-Luna variants, as well as older GPT-5.5 and GPT-5.2 entries. According to discussion on Hacker News, where the change drew more than 350 upvotes, the prior working limit stood at approximately 372,000 tokens. Some models retain a higher `max_context_window` of up to one million tokens, but the default window that clients actually request now caps at 272,000.
That distinction matters. The maximum context window represents what the model could theoretically accept, while the working `context_window` is what the Codex client is configured to send by default. In practice, that means developers who were previously feeding the agent hundreds of thousands of tokens of source code will now hit the ceiling sooner.
Why context length matters for coding agents
For a coding agent like Codex, context window is not a vanity metric. It determines how much of a repository the model can "see" at once when reasoning about a bug, generating a feature, or refactoring a module. A 100,000-token drop means roughly 75,000 fewer words of code, comments, and documentation loaded into a single prompt.
In real-world terms, that can be the difference between an agent understanding an entire microservice end-to-end and needing to operate with a partial, truncated view. Developers working on large monorepos or performing sweeping cross-file migrations are the most affected, because their tasks depend on holding many files in context simultaneously.
The change arrives as agentic coding tools compete aggressively on how much context they can handle. Rivals have marketed enormous context windows as a key selling point, framing the ability to ingest whole codebases as the path toward genuine autonomous software engineering. OpenAI's pullback narrows that gap at the default level.
Cost, quality, and the economics of long context
The likely motivation is cost. Longer contexts are dramatically more expensive to process, both in compute and latency, because the model must attend over every token in the input. High "needle in a haystack" failure rates also tend to creep in as contexts stretch, meaning models sometimes miss information buried deep in a long prompt. Capping the default window can reduce these accuracy issues while trimming inference spend.
OpenAI did not publish a separate announcement explaining the reduction. The change traveled only through the metadata refresh and the backport to the 0.144 release line, which is the branch shipped to most non-alpha Codex users. That low-profile rollout is itself notable: a decision that directly affects developer productivity was communicated through a code diff rather than a blog post or changelog.
For teams, the practical implication is straightforward. Tasks that previously fit within a single context pass may now require splitting work into smaller chunks, retrieving relevant files more selectively, or accepting that the agent has a narrower field of view. Some developers may route around the limit by relying on the higher `max_context_window` where available, though doing so typically requires explicit configuration.
A broader pattern of quiet adjustments
The Codex context cut fits a wider industry pattern in which AI labs tune the parameters of their models between headline launches. Window sizes, rate limits, and default behaviors are frequently adjusted through infrastructure updates that never reach a press release. Developers who depend on these systems increasingly watch the underlying repositories and API responses for early signals.
OpenAI's move also highlights the gap between marketing and reality. A model may support a headline context length, but the effective window that products actually expose can be considerably smaller, shaped by cost decisions that providers make on behalf of users.
Stay Ahead of AI
Model parameters shift constantly, and the tools developers depend on can change overnight. For timely reporting on releases, capability changes, and the economics of frontier AI, follow the breaking AI news at AI Buzz Wire.
Read more AI news →
