OpenAI Quietly Slashed Codex's Context Window by 27% — Check Your No-Code AI Workflows

> TL;DR: On 18 July, OpenAI merged a one-line JSON change into Codex that cut the reported context window for GPT-5.6 from 372K to 272K tokens, a 27% reduction. If your no-code agent, RAG pipeline, or automation assumes 372K, you are now silently losing context or paying 2x for input tokens once you cross the 272K line. Check your workflows this week.

On Saturday, while most of us were doing weekend things, OpenAI shipped Codex 0.144.6. Buried in PR #33972, a one-line JSON change titled "Backport refreshed bundled model metadata," was a correction that reduced the reported context window for GPT-5.6 Sol, Terra, and Luna from 372,000 tokens to 272,000.

That's 100,000 tokens. Gone. A 27% cut.

The Hacker News thread hit 360 points within hours. Developers who'd built their workflows around that 372K figure woke up to find their assumptions invalidated by what OpenAI is calling a "correction." The raw model spec hasn't changed, it's still 1.05M tokens, but the catalog metadata that Codex uses to manage sessions now reports 272K.

For no-code builders, this isn't developer drama to rubberneck. If you're using Codex via the API inside Zapier, Make, n8n, or any other no-code platform, you may already be losing context you thought you had.

What actually changed?

The short version: very little, and also quite a lot.

The raw GPT-5.6 model still supports a 1.05M-token context window with 128,000 max output tokens. That hasn't moved. What Codex 0.144.6 changed is what the client *reports* and manages as the default context budget. The previous metadata showed a 372,000-token window with a 95% effective-context multiplier, giving you about 353,400 usable tokens. The new metadata reports 272,000, and the auto-compaction threshold fires at 95% of that, so your practical ceiling before things get compressed together is around 258,000 tokens.

But here's the bit that matters more than the number itself. OpenAI's pricing documentation for GPT-5.6 states that prompts exceeding 272K input tokens are billed at 2x input and 1.5x output for the entire session. Not just the overage. The full session.

The old 372K figure meant roughly 81,400 tokens sat in that premium pricing band before you'd get any warning. The correction to 272K closes that gap, which is almost certainly the real reason this happened, but it also means any workflow built assuming 372K is now operating with substantially less room before either compaction kicks in or your bill doubles.

This is a billing boundary dressed up as a spec correction, and nobody should be confused about what it really is.

Who actually gets burned by this?

The HN thread is full of developers fuming about compaction losing detail mid-session. The people I'm more worried about are no-code builders who don't even know this happened.

Long-context agent sessions. If you've built an AI agent that maintains conversation state across multiple turns, processing a long document, iterating on a complex data analysis, debugging a multi-file project, you were probably counting on 350K+ tokens of usable space. You now have roughly 258K before auto-compaction fires. The HN comments are littered with developers saying compaction destroys enough detail that they'd rather start fresh. That's a terrible experience if your agent is customer-facing.

RAG pipelines with large document sets. If your no-code RAG setup, say a Zapier workflow that loads PDFs into a Codex-powered chat, was tuned to feed 300K tokens of context into each prompt, those prompts now cross the 272K pricing tripwire. You're not just getting less context; you might be paying double without realising.

Multi-file coding sessions. For anyone using Codex as the backend for a no-code coding assistant, via Bubble, Retool, or a custom n8n workflow, sessions that load several large files for context are now hitting the compaction wall 27% sooner. Multiple people in the HN thread said they've already moved to Claude Code specifically because of this: it gives you 1M tokens of context with compaction you can disable entirely. One commenter put it bluntly: "npm uninstall -g @openai/codex && curl -fsSL https://claude.ai/install.sh | bash."

Platform connectors that hard-coded 372K. Any automation platform or community template that baked the 372K figure into its Codex connector configuration is now silently wrong. If Make's Codex module or a popular n8n template assumed 372K, users are getting compacted earlier than expected and may not notice until something breaks subtly in production.

Can I check if my workflows are affected?

Yes, and you should do it this week.

Check your Codex API usage logs for any session that regularly exceeds 250K input tokens. If your workflows were comfortably sitting in the 300K to 350K range, you're now crossing the 272K pricing boundary. Look at your last billing statement: if you see sessions being charged at the premium tier without you explicitly opting in, that's the tripwire in action.

If you're running Codex locally or through a self-hosted harness, set `model_auto_compact_token_limit = 270000` explicitly. This postpones compaction to just under the pricing boundary. It's the workaround the community landed on before this PR even shipped.

For no-code platforms where you don't control the underlying config, Zapier, Make, n8n cloud, check whether their Codex integrations expose any context or compaction settings. Most don't. If yours doesn't, your best move is to audit how much context your workflows actually consume and either chunk your inputs more aggressively or switch to a model with a higher practical ceiling.

What's the bigger lesson here?

This is the third time in as many weeks that an AI dependency has shifted under no-code builders' feet. DeepSeek discontinues its legacy `deepseek-chat` and `deepseek-reasoner` API endpoints on 24 July. GitHub Models shuts down entirely on 30 July. And now OpenAI takes 100K tokens off the table with a one-line merge.

The pattern is hard to miss: AI model capabilities are not a contract. They're a moving target, and the direction of movement is not always "more." Sometimes it's "less, but presented as a correction" or "gone, with six weeks' notice."

I've been saying this in various forms across the last few editions of nocode.tech, but it's worth putting plainly: if your no-code product depends on a specific model's specific behaviour at a specific context window, you've built on sand. The platforms are optimising for their own costs and their own scaling challenges. Your 372K assumption was never in their calculus.

The builders who thrive here are the ones who build model-agnostic pipelines from day one. Abstract the AI call behind your own API endpoint. Keep model selection configurable, not hard-coded. Test your workflows against the pricing tripwires, not just the spec sheet maximums. And never assume that a capability that exists today will exist tomorrow at the same price.

The takeaway

Codex didn't technically lose capabilities; the underlying model is unchanged at 1.05M tokens. But the *effective* context you get before hitting either compaction or a 2x price multiplier just shrank by 27%. For most no-code builders, that distinction is academic: your workflow either works or it doesn't.

Check your API usage this week. If you're crossing 272K, you're either paying more than you think or losing context you're counting on. Fix it now, before it breaks something in production.

And take the broader lesson seriously. Between this, DeepSeek's API deprecation, and GitHub Models disappearing entirely, July 2026 is turning into a masterclass in why your no-code stack needs to treat AI model dependencies as temporary, not infrastructure.