Thirteen models now ship context windows over a million tokens. Llama 4 Scout advertises ten million. The number on the spec sheet has quietly stopped meaning anything useful, and teams keep designing systems around it anyway.

The 2026 long-context benchmarks all tell the same story. RULER, MRCR v2 and NoLiMa show accuracy holding for a while, then falling off a cliff well before the advertised ceiling. For multi-fact retrieval past 200K tokens, the gap between advertised and effective runs 30 to 60 points. Chroma tested 18 frontier models and found accuracy dropping more than 30% for facts sitting in the middle of the window. The rough rule people keep landing on is that models break 30 to 40% before their claimed limit. A 200K model gets unreliable somewhere around 130K.

The number on the box is a ceiling, not a working range

A context limit tells you where the API starts returning errors. It says nothing about where your accuracy starts sliding. Those are two different numbers and only one of them is published.

What actually happens as you fill the window is gradual. Retrieval of a single obvious fact stays fine for a long time. Multi-fact reasoning goes first. Then anything that requires holding two conflicting instructions from different parts of the input. By the time you notice, the failures look like model quality problems rather than context problems, so teams go shopping for a better model instead of shortening their prompt.

Bigger windows did not remove the cliff. They moved it further out. That is still useful, but it changes the engineering question from "does it fit" to "where does it stop working."

Agents feel this before chat does

A chat turn is one shot. An agent loop accumulates. Every tool result, every file read, every retry gets appended, and after twenty steps you are deep into the degraded zone without a single call having failed.

This is why agent runs often go strange around step 15 or 20. Instructions from the system prompt are now 300K tokens behind the current position. The model is still technically "reading" them. It just weights them a lot less than the tool output it saw thirty seconds ago.

If your agent gets vaguer the longer it runs, that is usually context depth, not reasoning ability.

Find your own cliff

Nobody can hand you this number. It depends on your documents, your task and your accuracy bar. But measuring it takes an afternoon.

Take real content from your own system, not synthetic filler.
Plant a known fact at 10%, 25%, 50%, 75% and 90% depth in the input.
Run the same question at 20K, 50K, 100K, 200K, 400K and up.
Score exact-match retrieval and, separately, a task that needs two facts combined. The second one will fail first.
Find the length where accuracy drops below what your product can tolerate. That is your real limit.

Then treat it as a hard budget in code. When the loop approaches it, compact: summarise older tool results, drop resolved sub-tasks, re-inject the system instructions near the end of the input rather than trusting they survived from the top.

The practical takeaway

Buying a larger window is not a fix for a context problem. It buys you more room to make the same mistake at greater expense, since you pay for every token you stuff in there whether the model uses it well or not.

Measure the cliff. Budget below it. Compact deliberately instead of hoping the window absorbs everything.

We are here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you are looking to build something, get in contact with us today.