Differences in simulation outputs using random

I'm trying to find the expected number of time steps for an share price to fall from $25 to $18 by generating simulations. Analytically, it can be shown that the expected number of time steps required is 129.

The following 3 codes appear fundamentally the same to me, but gives different results.

The first returns an average of 100 time steps:

target = 18
support_level = 20
I = []
n = 50000

for i in range(n):
    stock_price = 25
    t = 0
    
    while stock_price != target:

        t += 1

        if stock_price < support_level:
            stock_price += random.choices([-1, 1], weights=[1/3, 2/3])[0]

        if stock_price == support_level:
            stock_price += random.choices([-1, 1], weights=[0.1, 0.9])[0]

        if stock_price > support_level:
            stock_price += random.choices([-1, 1], weights=[2/3, 1/3])[0]
    
    I.append(t)
sum(I) / len(I)

The second returns an average of 114:

for i in range(n):
    stock_price = 25
    t = 0
    
    while stock_price != target:
        x = random.uniform(0, 1)
        t += 1

        if stock_price < support_level:
            if x <= 2/3:
               stock_price += 1
            else:
               stock_price -= 1

        if stock_price == support_level:
            if x <= 0.9:
               stock_price += 1
            else:
               stock_price -= 1

        if stock_price > support_level:
            if x <= 1/3:
               stock_price += 1
            else:
               stock_price -= 1
    
    I.append(t)

The third returns an average of 129:

for i in range(n):
    p=25
    t=0

    while p != 18:
        if(p==20):
            dist = [0.1, 0.9]
        elif(p>20):
            dist = [2/3, 1/3]
        else:
            dist = [1/3, 2/3]

        p=p+random.choices(direction, dist)[0]
        t+=1
        
    I.append(t)
    
sum(I)/len(I)

What is the difference?