Grok 4.5 has just become a lot cheaper without anybody noticing or publishing anything, and it’s now the best “balanced” model to use for agentic tasks according to our internal benchmarks. Grok launched a few days ago with a price for input/cache/output of 2$ / 0,50$ / 6$. Looks cheap, but there’s a catch: the cache is very expensive, and it’s what makes up to 80-90% of used tokens in a real world agentic task. For reference, GPT Terra high’s pricing is 2,50$/ 0,25$ / 15$. Despite being apparently a much more expensive model, it's actually cheaper in real world tasks. However, when I run our Vibetasking benchmarks today, Grok was the cheapest model (even cheaper than GPT 5.6 Luna). Trying to understand why, I found out Grok 4.5 cache price is now 0,30$ and I see no trace on the internet of anybody saying anything about it. With this, Grok is a lot cheaper, while being extremely smart, becoming the best model by far to use as the reasonable, balanced option for most tasks, and really is on the Pareto Frontier now. This is soo good.
@ArtificialAnlysyou should re-run your tests for the model, I'm sure the price per task will decrease significantly.@kimmonismusFYI@Alezander907you'll need to re-run your tests :)