It's about 39%. The problem is that in a normal year, everyone comes out an exam with a rough sense of how they've done. If you come out feeling like you bombed it then seeing your B become a C is not going to be a surprise.
The issue is that "predicted grades" are not expectation values in a statistical sense, if they were then teachers would be pretty consistently doing a bad job of forecasting them.
A predicted grade is based on an assessment of ability: "What do I think little Johnny is capable of when he takes his exam in however months time". Now in theory, results can be both better and worse than that prediction. In practice, the fact that the scale is closed ended (and many more are near the top than the bottom) means that it is more common for the variability of exam performance to lead to slightly worse performance. In other words, more pupils underperform their ability based prediction on exam day than outperform it.
So there is nothing inherently wrong with reducing grades from the centre assessed grades, many newspaper headlines are treating this like the CAG is the "real" grade that is somehow being taken away.
If you look at this table: https://twitter.com/Samfr/status/1293979033458417668/photo/1 you can see that the overall grade distribution produced by the model for this year is similar to that for previous years. Not surprising since that is by design.
The problem here is that Ofqual is treating grade inflation like it's the worth possible thing that could happen and has essentially sacrificed individual fairness on the altar of aggregate consistency.
Ranking students might be roughly correct, but many of the rankings in the middle where students are very similar are basically random.
Imagine there are two schools: A typical small private school, which normally achieves high grades, and a larger state school which normally achieves a broader spread. Normally the highest individual grade is achieved at the private school.
The historic average of the private school might be A-star, A, B, but predictions were A-star, A, A. Rankings are correct to the individuals' abilities. "Actual" or "deserved" standard could be A, A, A or A, A, B.
The state school's top 3 grades are normally A, A, B. This year you have a student who's capable (and likely) to achieve an A*. Teacher predictions for the top three grades are therefore A-star, A, A. Rankings are still correct according to ability.
Across the country the top 6 grades are normally A-star, A, A, A, B, B.
The private school uses the teacher predictions, because it has a small class: A-star, A, A.
We're now left with only A, B, B available to give to the top students at the state school, so the A-star, A, A assessed students are awarded A, B, B. Every one is a downgrade.
But it's OK, because across the country there's no grade inflation.
However, teachers naturally want to motivate their best performers (what teacher would grade so that the best student in their class would get a B- ?) and school administrators have a practical incentive to inflate the grades. This means that subjective evaluation of speeches and projects (especially if you're comparing them within a single institution, not to what projects and speeches are considered good and bad in other institutions) and continuous real-time questioning won't lead to an evaluation that's useful to determine where a student stands in comparison to other students in other institutions, and furthermore, they are going to be optimistic. A relevant headline is "40% of A-level results were downgraded from teachers' predictions." - however, that seems normal and expected to me; do we have the equivalent numbers from last year, comparing the teacher's prodictions with the actual exam results?
Technically you might consider taking the "local evaluation" and make an adjustment for the institution, but this is exactly what they attempted to do this year in UK, and the whole original article is about the limitations of this approach.
Also, fraud and cheating is a problem. There are many approaches that have proven to result in somewhat effective remote teaching and learning, however, we don't have good solutions for effective remote evaluation (this is a problem that I personally experienced this spring in my work). For the pre-COVID remote education, the only thing that IMHO worked was a network of in-person proctoring centers supervising the testing process and identity of the people who are taking the test (otherwise people will get others to take the test for them, it's not a hypothetical issue) or, in certain cases, a privacy invasive and labor intensive online proctoring.
IMHO it makes all sense to separate instruction and teaching from final evaluation and certification, due to the unavoidable conflicts of interest and misleading results that happen otherwise. This is also what UK has chosen as its policy - but as they found out, this means that they can't have a meaningful evaluation if they skip the actual evaluation part this year.