Short answer
No. After the very first kilobytes the large, easy-to-see biases get smaller, but they never vanish, they are not monotonic, and nothing is known (and nothing seems likely) that would let you point to a single number N and say “from here on the bias is minimal”. “RC4-drop-N” therefore has no magic value of N that turns RC4 into an unbiased generator – it only trades big early weaknesses for smaller later ones.

Long answer

1. Why the early bytes are so bad
   • The Key‐Scheduling Algorithm (KSA) produces an RC4 state that is far from the uniform distribution of all 256! permutations.
   • The PRGA that follows is completely deterministic and invertible. With a random key you can think of the initial state as a random variable S₀ with a distribution D₀. Each PRGA step is a bijection F, so after t steps the state is Sₜ = Fᵗ(S₀). A bijection only re-labels the probabilities, it cannot bring the distribution closer to uniform; it merely moves the same biases around.
   • What changes is which statistical tests expose the bias. Very simple ones – “is the second output byte 0?” – work spectacularly well at the very beginning. After a few hundred steps they already fail, but more sophisticated correlations still work.

2. What has actually been measured
   • Mantin & Shamir (2001), Fluhrer–McGrew (2001), Paul & Maitra (2004), Sen (2009), AlFardan–Bernstein–Paterson (2013) and others published whole families of biases that occur for all t, not just the first few bytes.
   Example (Mantin’s “second‐order” bias):
   P[ Zₜ = Zₜ₊₁ + 1 ] = 1/256 + 1/2¹⁶ for every t ≥ 1
   i.e. about 1⁄65536 above random no matter how far you are in the keystream.
   • Other biases are periodic in t (period 256, 512, …) because of the way the index i wraps around. So if you plot the deviation |P[Zₜ=x]−1/256| you get a curve that goes sharply down in the first ≈1000 bytes, then wiggles forever around a plateau of roughly 2⁻¹⁴ … 2⁻¹⁶.

3. Monotonicity – does dropping one more byte always help?
   Definitely not. Take any concrete measure of bias – say
   Bₜ = max_x | P[Zₜ = x] − 1/256 |.
   In empirical studies Bₜ falls quickly at first, but once you are past a few thousand bytes it oscillates: sometimes Bₜ₊₁ < Bₜ, sometimes Bₜ₊₁ > Bₜ. (You can verify this with a few billion randomly-keyed streams and a spreadsheet). Because the PRGA is invertible you cannot prove a general inequality such as Bₜ₊₁ ≤ Bₜ.

4. Asymptotics – does RC4 ever become “perfectly mixed”?
   No proof is known and none is expected. In fact, because the state distribution Dₜ never approaches the uniform distribution, for every t there exists some distinguisher with advantage at least as big as the original one; it may just be very complicated to describe. Thus the bias never goes to zero; it only gets harder to exploit.

5. What this means for “RC4-drop-N”
   • Picking N = 256, 768 or 3072 removes the spectacular first-byte and second-byte biases and most of the known key-correlated ones (which is why those numbers were once recommended).
   • It does not remove all exploitable bias; AlFardan & Paterson demonstrated practical ciphertext-only attacks on TLS even with N = 1536.
   • Because no monotone decrease exists, there is no provably optimal N, and there is no point after which the bias is guaranteed minimal.

Conclusion
Dropping more bytes will usually make the simple biases smaller, but it cannot eliminate bias and it is not guaranteed even to reduce whatever global measure you choose. No finite N minimises all biases of RC4, and therefore “RC4-drop-N” cannot be relied upon for long-term security.