Fireworks claims 1,521 tok/s for DeepSeek V4.1 Flash
A vendor number on a waitlisted product, with DHH supplying the only reaction worth quoting so far.
Fireworks says DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on its Ultrafast service, and calls that 7x faster than Claude Opus 5.5. The demo framing is that it is fast enough to ship Tetris in 11 seconds. Ultrafast is waitlisted, so the number is the vendor's own and nobody outside Fireworks has posted a run against it.
Fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5.
DHH's reaction was the one that travelled.
What the claim actually says
Two things are bundled together here and they are not the same claim. One is raw decode speed, 1,521+ tok/s, which is a property of the serving stack and the model size. The other is the 7x comparison against Opus 5.5, which depends entirely on what Opus was doing at the time and on which Fireworks has published no methodology in the post. The Tetris line is a demo, not a benchmark, and it is worth reading it as one.
There is also no independent confirmation yet, because there cannot be. Access is behind a waitlist. Until engineers with their own codebases get a key, the honest summary is that a vendor announced a number.
Why speed is the argument now
The reason this lands at all is that a lot of the agent tooling discussion lately has been about loops rather than single answers. A fast cheap model is attractive as a subagent, the thing you spawn twenty of to grep, summarise, draft and throw away, where one-shot quality matters less than how many turns you can afford. At 1,521 tok/s the economics of that pattern look different from the economics at 200.
Which is roughly what DHH was pointing at. He did not endorse the benchmark, he noted the scale of the shift, and the implicit comparison is to how recently 200 tok/s was the ceiling people planned around.
For working engineers, none of this tells you whether V4.1 Flash is any good. Speed claims and capability claims are independent, and the only number on the table today is the speed one.
I remember when 200tps felt impossibly fast.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.