What is left at the top is not vocabulary. It is being able to produce one meaning in several versions and to choose the version that fits the room: the neutral one, the formal one, the blunt one, the one you use with people you know well. A speaker who has only one version of everything is understood perfectly and is still visibly operating in a second language.
This is why the last stretch feels different from everything before it. Up to here, progress was adding things you could not do. Here it is producing something you can already produce, differently, on purpose.
Eight things that remain at C1 and above
- register is the remaining work. At this level errors are rarely grammatical; they are choosing a formal word in a casual sentence or the reverse.
- idiom and connotation. Two words with the same dictionary meaning carry different attitudes, and that difference is invisible in a dictionary.
- humour is the last thing to arrive, because it depends on shared reference and on timing rather than on vocabulary.
- reading widely, not more. Range across genres does more at this level than volume in one.
- write and be edited. Written correction from a skilled reader is the highest value feedback available above B2.
- accent is optional, precision is not. A strong accent with precise word choice reads as competent; the reverse does not.
- choose a domain to be excellent in rather than trying to be uniformly advanced, which is what native speakers actually are.
- the last ten per cent takes years, and knowing that in advance is what stops it being read as failure.
Three things, and none of them is range
Register control is the first and it is the largest. The same request exists as I would be grateful if you could, could you, can you and give me, and all four are correct English. Choosing wrongly is not a mistake anybody will name, and it decides how the sentence lands.
Precision in what stands beside a word is the second. At this level the wrong choice is rarely wrong, it is merely not what anybody says, and it is the fault a native ear catches fastest in otherwise excellent English.
The third is brevity, and it is the hardest. Advanced learners write and speak long, because length is where the range shows and because a longer sentence gives more chances to demonstrate. English at its most fluent is short, and a speaker who can say a complicated thing in eight words is doing something a speaker who takes thirty cannot.
Underneath all three sits understatement, which is a genuine barrier and is rarely taught. A great deal of English opinion is delivered below its actual strength: not ideal, a bit of a problem, I'm not sure that's right. Reading those at face value is one of the last things to go, and producing them is later still.
Where it stops: collecting rare words
The stall is the only kind of progress left that can be measured. Vocabulary size is countable, so an advanced learner keeps adding words, and the words available to add at that point are by definition rare ones.
Rare words are the least useful thing at this level. They come out in the wrong register, they mark the speaker as somebody who learned from lists, and they do nothing at all about the three things above. Meanwhile the learner feels productive, because the count went up.
The reason the stall holds is that the useful work is unmeasurable. There is no number for register control, no test that reports it, and no session where you can see it improve. It moves over months and it is only visible in how people respond to you, which is the least legible signal there is.
Getting four versions and being told what changed
Ask a model to rewrite one sentence of yours at four levels of formality and to state, for each version, which words carry the change. The second half of that instruction is the useful half: four versions on their own are a menu, and naming the mechanism is what transfers to the next sentence.
It replaced watching for the difference across a decade of reading and hoping to notice, which is how anybody used to acquire this and which is why it took a decade. Free at the level this needs.
The limitation is important and specific. It will produce four confident versions and it cannot tell you which one the people you actually work with use, because register is a fact about a particular community rather than about English. Take the mechanism and check the choice against the room.
What it looks like when it goes wrong
The recognisable failure is uniform formality. Everything the speaker produces sits at the same slightly elevated level, which is safe, which is correct, and which reads as distance in situations where distance was not intended.
The second is a rare word landing in a casual sentence. It is not an error and it is noticed by everyone in the room, and it is the clearest possible signal of a vocabulary assembled from lists rather than met in use.
The third is missing understatement in the other direction. Hearing that's not ideal as a mild remark, when it was the strongest objection the speaker was prepared to make out loud, produces a genuine misreading of a meeting.
The pairings that decide the second of the three is which words English puts together, and which it never does, and the far end of the register range, where being placed by a word is the whole risk, is what people actually say, and when not to say it.
What none of this is measured by is an exam, since the tests that report a number stop distinguishing usefully near the top of their scale, so the evidence at this level has to come from somewhere else: placing yourself honestly without paying for an exam.