For what it is worth, xAI models have historically always underperformed my personal, amateurish benchmarks when compared to what I’d expect to see due to their

For what it is worth, xAI models have historically always underperformed my personal, amateurish benchmarks when compared to what I’d expect to see due to their performance in more publicly known evals. Claude has historically always been the exact opposite, underperforming public evals while outperforming my testing and OAI with 4.1 and 5 output I found to most closely be within what one would expect from those same benchmarks vs my own. Training to the test, especially in flashy, easy to grasp things like bouncing ball and subsequent failure of similar tasks after small prompts changes made me doubtful of any claims that XAI models are good, with Grok 4 Fast not being an exception. Take the benchmark performance, even when they let external partners like Artificial Analysis do the evaluation, of any LLM lab with massive skepticism, but with XAI I have the most reason to doubt their claims due to prior performance discrepancies. Just to add some context to what I perceive is their honesty. Also, lest we forget their commitment to publishing system prompts and changes to that for the X deployment of Grok and how these somehow did not reflect any changes during the South Africa and Mechahitler incidents…