Aren't those speeds for the first few tokens, that, because of no context for attention to attend, are much faster to compute that the others? I expect the actual token speed to nosedive sharply as you get more context utilization.

What about the system prompt you're using? General purpose harness like Claude Code will insert their happy 22k tokens, even before your first useful token is processed. That might make it a no-go even _before_ you can even start, as the maximum context for this seems pretty limited (to make it fast)

And those LLMs are all "thinking", that is, rather that "one-shoting" the answer, they generate a lot of internal use reasoning tokens before starting to generate useful, visible response tokens. You can easily get to 30k tokens when your initial prompt is vague ambiguous garbage (as are naive transcriptions) as your LLM will "But wait, the user might have meant X, let me think more about this" lots of times.

No thinking (therefore much worse answers) will be a requirement.