A coder opens a project and everything looks normal.
The files load, the session starts. Then, midway through a multi-file rewrite, the tool hits a wall and the errors start stacking. No warning arrived. No update landed. Yet the tool that handled the whole codebase yesterday cannot finish the job today.
The question is not what broke, but what was quietly moved overnight.
How an invisible server-side ceiling can shrink a paid tool overnight
Context window is the working memory of an AI coding assistant: the total amount of code, conversation history and instructions it can hold in view during a single session. Think of it as a desk. The bigger the desk, the more files you can spread across it without losing track of where anything sits.
The critical detail most users never read is that this limit lives on a server, not in the app on their machine. That means the company can adjust the ceiling at any time, and the app on your computer has no way to warn you because the app itself has no idea it happened.
Once the window fills, earlier content is dropped or compressed. For a developer working across thirty or forty connected files, that forgetting does not look like forgetfulness. It looks like a crash, a corruption, or a tool that has simply stopped working.
The exact moment developers noticed the ceiling had dropped
On July 13, a GitHub issue reported that GPT-5.6 Sol’s effective context window inside Codex had dropped from roughly 353,000 tokens to roughly 258,000 tokens, despite the underlying model being advertised at a full 1.05 million tokens. Their multi-file refactoring tasks had started failing with context-overflow errors where they had not before. The issue was filed against CLI version 0.144.3 by paid ChatGPT-authenticated users across multiple environments.
No announcement. No changelog entry. No documentation update. The change was not announced; users discovered it by measuring where their sessions truncated. The model’s own published spec sheet still advertised a number that dwarfs both the old ceiling and the new one.
That is a 27 percent cut, felt immediately by anyone working on a project large enough to use what the tool had previously offered.
Where the verified numbers come from and what the model card actually says
The tool in question is the Codex command-line agent, which runs inside a developer’s terminal and edits files directly. GPT-5.6 Sol’s model card specifies a 1,050,000-token context window, but the Codex CLI model catalog exposes context_window: 272000 and max_context_window: 272000. Codex CLI version 0.144.6, released on July 18, updated the bundled model instructions and context windows. The raw model configuration diff shows the change as 372,000 tokens reduced to 272,000; user measurements of effective session limits put the drop from roughly 353,000 tokens to roughly 258,000.
The gap between what the model card says and what the tool actually permits is not small. At the standard 95 percent effective-context multiplier, the 272,000-token catalog cap yields roughly 258,400 usable tokens, around 25 percent of the advertised capacity. That distinction matters enormously to anyone who paid for a subscription expecting the published capability.
Because the Codex CLI is open source, the change surfaced as a GitHub issue rather than an announcement. Users filing the issue argued that any product-specific cap must be prominently documented before users select the model, and that context limits must not be silently reduced for existing or new paid-user workflows. Developers called this a “severe regression,” pointing to broken production workflows and automated pipelines built against the previous limit.
Why the new ceiling sits at exactly that number and not any other
Here is where the story deepens. There is a detail buried in the API documentation that makes the 272K number look less like a specification and more like a billing boundary.
Under OpenAI’s pricing, once input to GPT-5.6 Sol exceeds 272,000 tokens, the entire request is billed at 2× input and 1.5× output. The higher rate applies to all tokens in that request, not only the portion above 272,000. The 272K figure aligns with exactly that pricing threshold documented in OpenAI’s rate card, meaning the context cap functions as a billing guardrail, not a model limitation. Users cannot access the long-context pricing tier through Codex even if they wanted to pay for it.
Whether that alignment is deliberate cost management or coincidence, the practical outcome is identical: paying subscribers lose access to a wider window, silently. For anyone who wired automated agents or developer tools against the previous window, there is no changelog to explain what changed when the session fails at 2 a.m.
What the pattern reveals and where it leaves everyday users
Silent server-side adjustments are not unique to this one tool. The 272K limit was the ceiling during the GPT-5.5 generation, was raised to 372K when GPT-5.6 Sol arrived, and was then quietly returned to its earlier value within weeks, an oscillation that fed significant distrust among users who had built workflows against the wider window. The pattern suggests OpenAI is actively tuning the cost-quality tradeoff on the server side, an operational decision that would benefit enormously from being communicated proactively rather than discovered through broken sessions.
The fix most engineers recommend is to never hardcode a limit into any workflow and to validate what the server is actually offering at the start of each session. That is solid advice, but it puts the burden on the buyer to audit a product they have already paid for.
The honest takeaway for anyone using an AI coding tool right now: the number on the marketing page describes the underlying model’s ceiling, not necessarily what any given product serves. That gap between the two is the real delivered spec. For users who followed the push toward AI-assisted workflows across every corner of their workday, this is a useful reminder that advertised capability and delivered capability can diverge without warning, without an email, and without a single line in a changelog.
