Codex adjusts GPT-5.6 context: what changes for long sessions

On July 18, 2026, OpenAI merged a hotfix for Codex’s 0.144 branch that keeps a 272,000-token context window for GPT-5.6 Sol, Terra, and Luna. It is a technical detail that can go unnoticed in a short request, but it matters when a coding agent works for hours, examines many files, or chains together tool results. In those situations, context is the working memory that lets the system connect what it has just read with decisions made earlier.

The update does not, by itself, change GPT-5.6’s underlying capabilities. It affects how Codex configures the information space it keeps available before compacting part of a conversation. That distinction matters: a model may have a context capability advertised in an API while a particular product reserves, manages, or limits usable context differently in its own sessions.

What a context window is

A context window gathers the material a model can consider in an interaction: instructions, code snippets, conversation history, terminal output, and tool responses. In AI-assisted programming, that memory can grow quickly. An agent inspecting a repository does not merely read files; it also receives directory listings, build errors, test results, diffs, and decisions that change the direction of the task.

As the limit approaches, Codex can compact older information to make room for new material. Compaction does not necessarily mean erasing everything: it can retain a summary of earlier work. But a summary does not hold every detail of a conversation, which is why longer tasks may need to reopen files or reconstruct earlier decisions. The practical effect depends on project size, the type of work, and how a session is organized.

The confirmed change, and what remains unknown

OpenAI’s public pull request is titled “Narrow 0.144 hotfix to GPT-5.6 prompts and context.” Its summary says that it keeps 272,000 tokens for both context and maximum context windows across the three GPT-5.6 variants. It also says the hotfix retains refreshed prompts and corrected context metadata. The text does not provide a public explanation for choosing that figure, nor does it detail an expected effect on speed, cost, or quality.

InfoWorld reported on July 20 that the default configuration had moved from 372,000 to 272,000 tokens and quoted criticism from some developers. Those reactions are worth noting because they come from users of long sessions and large repositories, but they are not an independent measurement of a broad performance decline. They also do not establish that every coding task will be affected in the same way.

OpenAI introduced GPT-5.6 on July 9 as a three-tier family—Sol, Terra, and Luna—available in Codex depending on a user’s plan. Its official help page also notes that usage limits depend on plan and system conditions. The documentation confirms that product experience is not reduced to one number: client version, access level, tool use, and service policies all contribute to what a developer sees.

Working when context is limited

For a focused bug fix, 272,000 tokens may be more than enough. The challenge appears when a task must keep a detailed picture of a large codebase across many iterations. In that setting, it helps to split the goal into verifiable deliverables, record decisions and invariants in project files, and ask the agent to identify which files it examined before changing others. That is not a magic fix; it reduces dependence on one very long conversation.

Tests, human review, and version control also help. Compaction can change which details are present in a response; external validation keeps an independent reference beyond the agent’s own summary. For teams, the lesson is not to avoid coding agents, but to design workflows that do not depend on one operating limit staying fixed.

The verifiable fact is narrow: Codex 0.144 keeps 272,000 tokens of context for GPT-5.6 Sol, Terra, and Luna. What is missing is a public OpenAI explanation of the decision and data measuring its effect. Until then, developer criticism is a useful signal about long-running work, not a final verdict on GPT-5.6.