There is at least 6 orders of magnitude difference in computation cost of computing a token (pass through a multi-B model) and doing a single step of a program.

There is at least 6 orders of magnitude difference in computation cost of computing a token (pass through a multi-B model) and doing a single step of a program. Even if your validation is really naive, it's hard to beat 6 orders of magnitude. So no, you're not generating tokens fast enough.