I would guess that it is essentially just a “grok 4 mini”, but if you use mini as the qualifier then most people will be inclined not to use it. If you call it fast then it gives people a reason to select it.

Currently ~160 tps, per https://openrouter.ai/x-ai/grok-4-fast:free

They sound like they’re positioning it more that it’s faster to complete because it uses fewer tokens - see the mentions of token efficiency.

Ahh so this might be the Sonoma sky Alpha that was gathering feedback on openrouter recently.

I tried that one extensively (it was free) and was disappointed vs regular grok 4 so also maybe not.

It’s a little dumb but in my own use better than somnet

I've found the results shift quite a lot between models and updates. Deepseek is pretty consistently good at writing code that is rather easy to improve from mid to good quality. Claude used to be pretty good, but now writes 10x the code you'd need. Gemini is amazing, if you buy one of the more expensive tiers, which in turn isn't really worth it because there are so many other options. GPT and Grok are hit and miss. They deliver great code or they deliver horrible code. GPT and Claude have become such a hurdle I've had to turn github co-pilot off in my VScode. Basically I use deepseek for brainstorming and GPT for writing configs, queries, sql and so on. If either of them fails me I'll branch out, and Grok will be on that list. When I once in a while face a real issue where I'm unsure about the engineering aspects, I'll use one of my sparse free gemini pro queries. I'd argue that we should pay for it at my work, but since it's Google that will never happen.

From an ethical perspective, and I'm based in Denmark mind you, they are all equally horrible in my opinion. I can see why anyone in the anglo-saxon world would be opposed to Elon's, but from my perspective he's just another oligarch. The only thing which sets him appart from other tech oligarchs is that he's foolish enough to voice the opinion publicly. If you're based in the US or in any form of Government position then I can see why DeepSeek is problematic, but at least China hasn't threatened taking Greenland by force. Also, where I work, China has produced basically all of our hardware with possible hardware back-doors in around 70% of our IOT devices.

I will give a shoutout to French Mistral, but the truth is that it's just not as good as it's competition.

Really, you are "surprised" to see the negativity here?

How do you manage sending and receiving requests to multiple LLMs? Are you going it manually through multiple UIs or using some app which integrates with multiple APIs?

I believe, despite all the hate it got today, we'll one day be grateful that there is at least one big AI provider chooses a route with less lobotomy.

It doesn't outperform uniformly across benchmarks. It's worse than Grok 4 on GPQA Diamond and HLE (Humanity's Last Exam) without tools, both of which require the model to have memorized a large number of facts. Large (and thus slow) models typically do better on these.

The other benchmarks focus on reasoning and tool use, so the model doesn't need to have memorized quite so many facts, it just needs to be able to transform them from one representation to another. (E.g. user question to search tool call; list of search results to concise answer.) Larger models should in theory also be better at that, but you need to train them for those specific tasks first.

So I don't think they simply trained on the benchmark tests, but they shifted their training mix to emphasize particular tasks more, and now in the announcement they highlight benchmarks that test those tasks and where their model performs better.

You could also write an anti-announcement by picking a few more fact recall benchmarks and highlighting that it does worse at those. (I assume.)

> Can anyone explain why that makes sense?

Can be anything from different arch, more data, RL, etc. It's probably RL. In recent months top tier labs seem to have "cracked" RL to a level not seen yet in open models, and by a large margin.

Just two different models branded under similar names. That's it. Grok 4 is not the slower version of Grok 4 Fast, just like gpt-4 is not the slower version of gpt-4o.

Grok 4 Fast is likely Grok 4 distilled down to remove noise that rarely if ever gets activated in production. Then you'd expect these results, as it's really the same logic copied from the big model, but more focused.

If this is sonoma-dusk that was on preview on openrouter, it's pretty cool. I've tested it with some code reverse engineering tasks, and it is at or above gpt5-mini level, while being faster. Works well till about 110-130k tokens tasks, then it gets the case of "getthereitis" and finishes the task even if not all constraints are met (i.e. will say I've solved x/400 tests, the rest can be done later)