Hacker News new | ask | show | jobs
by mythz 43 days ago
Ok speed (202.7 tok/s) and value (1.25 -> 2.50) look great, with pretty decent intelligence.
4 comments

The problem with speed is that they usually are very fast for first few weeks and then suddenly much slower. They did such trick when they advertised Grok 4 fast ( dropped from 200 tps to 60tps)
Grok 4.1 is still 110tps. The only other model that comes close is Gemini at 85tps.
Wow. That is a big drop.
202.7 tok/s is only OK speed? Which providers are you using that are significantly better than that?
for reference, it's the 2nd fastest model tracked in the "Highlights" section of https://artificialanalysis.ai/
Yes, it's incredibly fast. Openrouter is clocking 60 tokens per second, which is on par with the likes of sonnet, opus, GPT 5.5.
That section misses Cerebras and Groq which are up to 5x faster.
Very different tech and limitations though so wouldn’t make sense to compare 1:1 I think
What are the limitations ?
Much smaller context
I said speed was great, Cerebas and Groq can provide better performance, likewise Fast versions of Cursor's Composer and Claude.

The reported speed like benchmarks is only a reported number on paper, we'll see how it holds up in real world usage, so far OpenRouter is only reporting 73tps

[1] https://openrouter.ai/x-ai/grok-4.3

i really don't trust openrouter numbers.

i use byok and see responses fail on openrouter while they work perfectly at the provider. the provider is often listed as 'down' and it's very clearly up on the original api and serving requests.

cerebras quotes oss 120b at 3000tps and it is under 800 on openrouter.

same with fireworks, i am getting much higher numbers not on openrouter. but recently i think fireworks deepseek is kind of spotty, the main provider i know that just doesn't go down is vertex and they charge 2-3x the rest

Their stats look ok, but when I tested it[0], it was 4x slower than 4.20.

[0]: https://aibenchy.com/compare/x-ai-grok-4-20-medium/x-ai-grok...

Value should be calculated some other way, like cost per task completion or something.