Rice's theorem doesn't apply because your language isn't semantic.
Let's call your language $L$, and assume that both $M_1$ and $M_2$ accept $\Sigma^*$ for some input alphabet $\Sigma$.
$M_1$ directly accepts all inputs, but $M_2$ always reads the whole string and then accepts.
Then $L(M_1) = L(M_2)$ but $\langle M_1 \rangle \in L$ whereas $\langle M_2 \rangle \notin L$.
Remember, that
$$A_\varepsilon = \{\langle M \rangle : M \text{ is a TM that accepts } \varepsilon\}$$
is undecidable.
We now try to show that $A_\varepsilon \leq_m L$, from which follows that $L$ must be undecidable as well.
Proof:
For a TM $M = (Q, \Sigma, \Gamma, \delta, q_0, F)$ define $M' = (Q, \Sigma', \Gamma', \delta', q_0, F)$, where
- $\Sigma' \cap (\Sigma \cup \Gamma) = \emptyset$,
- $\Gamma' := \Sigma' \cup \Sigma \cup \Gamma$, and
- $\delta'(q, a) = \begin{cases}
\delta(q, \square) & \text{if } a \in \Sigma' \\
\delta(q, a) & \text{otherwise}
\end{cases}$, where $\square \in \Gamma$ denotes the blank symbol.
$M'$ essentially behaves like $M$, except that it ignores all of it's input characters, it treats them like $\square$.
So the computation of $M'$ on any input is basically the same as $M$ on $\varepsilon$.
Now, if $M$ accepts $\varepsilon$ in $n$ steps, then $M'$ accepts every word $x \in \Sigma'^{n + 1}$ in $n$ steps as well.
If $M$ doesn't accept $\varepsilon$, then $M$ accepts nothing.
Therefore
$$\langle M \rangle \in A_\varepsilon \iff \langle M' \rangle \in L$$
which implies that $A_\varepsilon \leq_m L$.