Agent Wars!

Nathaniel Whittemore (0:00)

And just like that, the AI agent wars have begun.

Meta's Muse personal agent has been a breakout consumer success. This week, the app surged over ChatGPT to be the number one free app in the US, and there are reports from satisfied users all over social media. But with that sort of success brings competition. And this weekend, Amazon decided to cut off Muse's ability to shop on Amazon sites. Will that impact Muse's momentum? Does agentic shopping even matter? As personal agents become a thing, we have a whole new set of questions to explore.

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around 5 minutes. SpaceX AI has kicked off what could be a big week for model releases with the launch of Grok 4.7. They called the model a notable improvement over Grok 4.6 at the same price and speed. Now, Grok models in general are competing in the increasingly difficult middle ground between ultra-cheap and cutting edge frontier. Grok 4.6 lagged behind GPT-56 Sol and Fable 51 on performance and was outcompeted on cost by Muse 1.2. Still, the model had its fans and was clearly capable of driving the success of Grokbot. And what's more for many people, showed that SpaceX AI was very much not out of the model race. This release sees some significant improvements, at least on the benchmarks. For coding, Grok 4.7 picked up 6 points on Cursor Bench 4 to overtake GPT-56 Sol, but is still 5 points short of Fable 5.1 score. On DeepSwee, the model improved by 6 points to overtake Fable 5.1, coming in just short of GPT-56 Sol. Purely on the benchmarks, then, Grok 4.7 looks like it should be a competitive coding model at a discount price. SpaceX highlighted significant improvements on long-horizon agentic work. For AA Briefcase, which measures multi-hour white-collar work, the model scored 1657 ELO points, putting it ahead of GPT-56 Sol and very close behind Fable 5.1. Benchmark scores for legal, electrical engineering and healthcare were similarly impressive. In a practical demonstration of the upgrades, SpaceX AI showed off a head-to-head comparison of an open game world. Grok 4.6's version was pretty low quality and unimpressive, while Grok 4.7 did a noticeably better job on both graphics and physics. Artificial Analysis gave the model a fairly favorable review, ranking it seventh on their intelligence index behind Astra, two iterations of Fable, Opus, MuSpark 1.3 and GPT-56 Sol. And on the coding agent index, it was ranked fourth inching ahead of GPT-56 Sol, but falling short of Opus, Astra and Fable 5.1.

Elon Musk celebrated the result, declaring that SpaceX AI is now the third-place lab for AI and decoding behind OpenAI and Anthropic. He wrote, When factoring in that Grok is significantly faster and lower cost, it's a great choice for your everyday workhorse.

Yet, as is sometimes the case, benchmarks appear not to tell the whole story, and the model's initial impressions didn't survive contact with public testing. Bobby showed his results in an AI rendering test against KimiK3, with both models tasked with animating a rocket takeoff. Grok's animation was fairly bizarre with the screen wobbling all over, leading Bobby to ask, what's wrong with Grok? Scott animated a star-shaped jello mold, and while he gave the render a passing grade, he noted it took 40 minutes, while Astra's version of the test took just five minutes. Theo declared that Grok's version of his fishslop game was the quote, worst I've seen this year.

A few people did have slightly more positive rendering results. OpenClaw maintainer Tac showed off a pretty slick animation made in Blender, although his prompt was just to make something cool that can be accomplished in 10 minutes. But of course, if you're sitting here thinking to yourself, are we really going to judge a new model based on 3D renders that are completely outside the use cases of most people? I think that's a fairly decent thing to ask. AI developer Kun Chen came to the defense of Grok commenting, ignore the reports that say it's terrible and the only thing they reference is a public benchmark. The same benchmarks told us Opus 5 was better than Fable. They are useless. Also ignore the reports that compare models with 3D games. That's not real work. It's made for attention on social media. I used Grok 4.7 for a whole day as my first mate and it has been a really solid model with visible improvements over 4.5. Cut noted, I'm ignoring 4.6 because 4.5 has been working better in my experience.

Still, overall, the first impression verdict on Twitter is not good. V posted, So let me get this straight. After releasing Grok 4.6 a month earlier, Elon spent 10 days teasing an imminent Grok 4.7 release, only to postpone it and then shift the narrative from pacing AI to vague claims about how 4.7 and 4.8 would be fable killers. Now, Grok 4.7 is finally out, and it's giving Timu Sonnet 5 vibes, believing benchmarks in September should be a crime. In some ways, I think the trajectory of the latest Grok models shows how difficult this middle space between state-of-the-art and really, really cheap actually is. AI entrepreneur and content creator Theo wrote, Grok 4.5 was an incredible model for the price, fast, pleasant to use, reliable, solid default model. Grok 4.6 was a, forgivable, step in the wrong direction in my opinion, slower and more expensive, using way more tokens per task for a slight edge in intelligence. I get it though, they have to climb benchmarks. Grok 4.7 is much harder to forgive. They claimed it would be more token efficient and it's less by 30 to 80 percent. It scores worse than Grok 4.6 in various benchmarks. It's slower, it's less pleasant to use, and real-world costs come out to more than 2x above Grok 4.6, putting it over Astra's costs in real-world use. This was a very disappointing release. I hope that the SpaceX AI team can acknowledge that and impress us with the next one.