A Codex context window is 272,000 tokens in one task

A 272K context window means Codex can work with approximately 272,000 tokens of information in one task. This can include your instructions, conversation, files and code. The larger the context window, the more background Codex can consider before older material may need to be compacted. In the Codex CLI, type /status to see the active model and remaining context for the session you are in.

The Codex context window for GPT-5.6 is configured at 272,000 tokens in the current stable Codex model metadata. InfoWorld reports that this is down from 372,000 tokens, a reduction of about 27% in the active working space available before older material may need to be compacted.

That does not mean GPT-5.6 suddenly became less intelligent. It means long coding or knowledge-work sessions may reach the point of context compression sooner. The practical impact depends on the size of the project, the length of the session, the tools in use and how well important decisions are stored outside the conversation.

Direct answer: The change matters most for large repositories, long-running agents and multi-stage work. Short, focused tasks are less likely to notice it. Keep important decisions in durable project files, break work into verifiable stages and do not treat one chat thread as the only record of the project.

What changed in the Codex context window?

On 18 July 2026, an OpenAI Codex pull request for the stable 0.144 release line said the bundled metadata would keep a 272,000-token context and maximum context window for GPT-5.6 Sol, Terra and Luna. The pull request described these as corrected context-window values.

InfoWorld reported that the configured GPT-5.6 input context in Codex moved from 372,000 to 272,000 tokens. The publication connected the change to concerns about more frequent compaction in large codebases and long-running sessions.

The difference is 100,000 tokens, or approximately 26.9%. That figure describes the change in configured working context reported for Codex. It is not a benchmark score, a model-intelligence score or a universal limit across every OpenAI product.

Why 272K and 1.05 million can both be correct

Context limits are product- and configuration-specific. OpenAI’s current documentation lists different figures for GPT-5.6 Sol depending on where it is used.

OpenAI’s GPT-5.6 Sol API page also says requests with more than 272K input tokens use higher token pricing. Its ChatGPT Business limits page lists 272K for GPT-5.6 Sol.

So the accurate question is not only, “What is GPT-5.6’s context window?” It is, “What context is available for this model in this product, on this plan, with this current configuration?”

What a context window is-and what it is not

A context window is the active working space a model can use for the current request or session. It may contain instructions, conversation history, code, documents, tool results and reasoning material needed to continue the task.

When that working space fills, an agent may need to compact, summarise, reload or drop older material to make room for new information. Good compaction can preserve the important thread of a project. It is still not identical to retaining every earlier detail verbatim.

A context window is not:

• the model’s general intelligence;
• its training knowledge;
• a guarantee that every token will receive equal attention;
• the same as long-term memory or project files;
• a substitute for version control, documentation or review.

This distinction matters because “smaller memory” is a useful headline but an imprecise technical description. A smaller configured context can affect continuity without proving a drop in reasoning quality.

How to check the Codex context window

In the Codex CLI, type /status in the current chat. OpenAI’s command reference says this shows the session configuration and token usage: the active model, the approval policy, the writable roots, and the remaining context capacity.

That figure is for the session you are in. It is not your account-wide token history. For account token activity, type /usage.

To keep the context figure on screen, type /statusline, turn on the context items, and confirm. OpenAI says the footer updates immediately and the choice is saved in config.toml under tui.status_line.

/status does not mean every OpenAI product uses 272,000 tokens. The Codex configuration and the API model page can list different limits. Check the product and plan you are actually using.

When the 272K Codex context window is most likely to matter

The reduction is most relevant when a task accumulates a large amount of active material, for example:

• repository-wide refactoring across many packages;
• migrations that span several services and deployment steps;
• long debugging sessions with repeated logs, screenshots and experiments;
• extended research that combines many documents and web sources;
• multi-day tasks where requirements and decisions evolve;
• agent workflows that call many tools and return large outputs.

A small bug fix, a short document edit or a focused analysis may never approach the effective limit. A broad project can reach it faster, especially when the conversation includes duplicated logs, large pasted files or repeated background explanations.

The real operational risk is not simply “running out of tokens.” It is losing a decision, constraint or piece of evidence that was never recorded anywhere else.

Six ways to protect long-running Codex work

1. Scope each task around one verifiable outcome

Give the agent a concrete objective, the relevant folder or repository, and a clear definition of done. A focused task reduces irrelevant context and makes review easier.

For a large programme, create a sequence of bounded tasks instead of one conversation that tries to hold the entire programme at once.

2. Store important decisions in durable project files

Write architecture decisions, product constraints, release gates and naming rules into the repository or project workspace. Use an AGENTS.md, project brief, decision log, runbook or handoff file where appropriate.

Conversation context is helpful. Durable files are inspectable, searchable and available to the next task even when the active window changes.

3. Use checkpoints before the work becomes difficult to reconstruct

Pause after meaningful stages to record what changed, what was verified, what remains open and which evidence supports the next step. A checkpoint should make it possible to continue without rereading an entire conversation.

OpenAI’s guidance for long-running Codex work similarly emphasises breaking ambitious goals into verifiable steps and preserving continuity across workstreams.

4. Keep tool output selective

Avoid pasting entire logs, repositories or exports when a bounded excerpt, file path or search result will do. Let Codex retrieve the relevant material from the real source when possible.

Selective context is not about hiding evidence. It is about keeping the evidence needed for the current decision visible and reducing duplication.

5. Commit and test at meaningful boundaries

Version control is external memory with accountability. Small, meaningful commits preserve known-good states, make diffs reviewable and reduce the cost of restarting a task.

Do not claim a commit, push or deployment happened unless it was verified. Context management works only when the durable record is real.

6. Start a fresh task when the objective changes

If a debugging task becomes a redesign, or a content draft becomes a publishing operation, create a clean handoff and start the new objective with the relevant facts. Carry forward the decisions and evidence, not every line of the old conversation.

What business teams should monitor next

The 272K configuration may change again. Model availability, product limits, pricing and agent behaviour can evolve independently. Teams using Codex for important work should monitor:

• the Codex version and model selected;
• documented context limits for the actual product and plan;
• how often compaction occurs in long tasks;
• whether important instructions survive checkpoints;
• repeated searches or work that may signal lost context;
• the quality of project files, handoffs and version history.

This is one reason resilient AI operations matter. Read our earlier briefing on AI operations changes business teams should act on for a wider view of model, platform and workflow changes.

Frequently asked questions about the Codex context window

Did GPT-5.6 become less intelligent?

There is no evidence in the cited update that GPT-5.6’s reasoning capability was reduced. The documented change concerns configured context metadata. A smaller active context can affect continuity in long sessions without being the same as lower intelligence.

Is the GPT-5.6 context window always 272K?

No. OpenAI currently lists 272K for GPT-5.6 Sol in ChatGPT Business and in the referenced Codex metadata, while the GPT-5.6 Sol API page lists a 1.05-million-token context window. Check the product, plan and current documentation.

Will ordinary Codex tasks be affected?

Short, well-scoped tasks may show little practical difference. Large repositories, long-running sessions and multi-stage work are more likely to encounter earlier context compaction.

What should I save outside the Codex conversation?

Save requirements, decisions, constraints, test results, source links, ownership, open risks and the exact next step. For software work, use project documentation and version control. For business work, use a durable brief, decision log or handoff.

How do I check the Codex context window?

In the Codex CLI, type /status. It shows the active model and remaining context for that session. Type /statusline to keep context stats in the footer. Type /usage for account token activity, which is a different figure.

Can SEO or AI search guarantee this page will rank?

No publisher can guarantee a ranking. Clear sourcing, direct answers, descriptive headings, structured data, useful internal links, crawlable pages and genuinely helpful content improve discoverability, but search engines and AI systems decide what to crawl, cite and rank.

The practical takeaway

The Codex context window change is a workflow signal, not a reason to abandon long-running AI work. Treat the active conversation as working space, not the only source of truth. Keep the objective focused, store decisions durably, verify in stages and preserve real project history.

Ankor Business Solutions builds practical, human-guided AI business systems for clearer operations. If your team needs a more durable way to organise decisions, responsibilities and AI-assisted workflows, start with the AI business tools overview, compare Ankor Nexa, or browse the current shop.