make 90% of your rivals at 20% of your price

we carry a pretty crazy week regarding AI model releases. The spotlight may have gone to the long-awaited GPT-5.6 family (with its Sol, Terra, Luna variants), but we have also seen how SpaceXAI Grok 4.5 was released and Meta too was renewed with Muse Spark 1.1.

In both cases we are facing a promising qualitative leap, and according to internal and independent benchmarks the performance of both models is already remarkably close to that of the frontier models from OpenAI and Anthropic.

Grok and Muse Spark get going

That was unthinkable a few months ago. Both Meta and xAI (now fully merged with SpaceX in the new SpaceXAI) seemed to play in the second (or third) division in the AI ​​​​segment.

SpaceXAI has barely shared results from Grok 4.5, but the table with Muse Spark 1.1 would look like this. Source: Meta (AI modified to include Grok).

In fact, Meta had been working on a mysterious new model called Muse Spark that we finally met in Aprilbut the result was lukewarm: it stood out in practically nothing and it seemed that the firm had not fully met expectations.

Something different happened with Grok. Musk’s ambition in this area has always been enormous, but from the beginning Grok’s approach was strange. That first version seemed to want to laugh at itself (and at all of us) with his sarcastic tonebut its evolution showed that we were facing a model with remarkable potential that I couldn’t quite find the key.

However, Both companies have given a more than promising push to these models. At SpaceXAI they have probably benefited from the recent acquisition of Cursor, which had its own background with Composer 2.5 and which has undoubtedly influenced the creation of a much more competent Grok 4.5 when it comes to programming. With Meta and its Muse Spark 1.1 the same has been seen: it is much closer to the great rivals, Fable 5 and GPT 5.6.

But the striking thing is not that they have come closer, but that they have done so by cutting the price significantly. And that is its great asset.

It’s not more for less, it’s almost the same for much less.

What is really relevant about these models is not only that the leap in quality has been significant: it is that Its cost is really competitive. In SpaceXAI’s announcement it is in fact something they especially highlight by saying that “Overall, Grok 4.5 offers the greatest intelligence per unit of time and cost.” At the launch of Meta there are no direct statements about this, but there is no need: it is the cheapest frontier model among its rivals, as can be seen in this table:

The question is, of course, whether those prices really pay off. If we get “enough intelligence” per dollar, to understand each other. And that is precisely what the Artificial Analysis ranking which evaluates the cost of completing predefined tasks in its benchmarks.

At the moment Muse Spark 1.1 is not included in that ranking, but for now there are two clear winners: on the one hand, GPT-5.6 Luna. On the other hand, even better, Grok 4.5.

Look at the green quadrant at the top left: the only models that fit are GPT-5.6 Luna and Grok 4.5. Muse Spark 1.1 is still untested. Source: Artificial Analysis.

That conclusion seems to make it clear that those specific SpaceXAI and OpenAI models are fantastic for intensive tasks where token consumption skyrockets. They may not be ideal If you are looking to solve very complex mathematical conjecturesbut for agentic tasks we have at least those two clear candidates. Muse Spark 1.1 probably meets those premises as well, judging by the benchmarks and cost.

We are therefore facing an interesting turn of events. One that also poses a real alternative to what was until now the great asset of Chinese open models: its cost was much lower than the frontier models of OpenAI or Anthropic, but performance was also lagging behind.

What Grok 4.5 and Muse Spark 1.1 (and GPT-5.6 Luna) offer is something like the best of both worlds: they manage to “give us” 90% of what the best models in the world do, but they do it at 10 or 20% of their price. Good, pretty and cheap.

Not a bad bet, of course.

In Xataka | A teacher suspected that his students had cheated with AI. He took the exam in person and his grades dropped by half.