Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive
I pay for three Codex subscriptions at $200 each, and for the past week they have mostly bought me waiting. Since I…
Perhaps I was too quick to dismiss Grok.
I'm not saying that Grok 4.5 suddenly beats everything else out there. It doesn't. If I had an unlimited budget and a really nasty coding task, I would probably still try Fable 5 or GPT-5.6 first. But now it is not enough to have the best model. Distribution is what xAI suddenly has.
A few months ago, the company seemed strangely irrelevant for serious coding work. Grok was there, but real developer workflow meant Claude Code, Codex, or Cursor. xAI had the big compute story and the usual internal chaos. TechCrunch even reported that all 11 xAI cofounders had left the company by the end of March. That is not the kind of setup that would have me waiting for a coding-model comeback.
And then SpaceX went for Cursor.
AP reports that SpaceX is moving forward with a $60 billion acquisition of Cursor. Cursor said in April it was teaming up with SpaceX because compute was their bottleneck and they wanted to use Colossus infrastructure.
In that environment, the Grok 4.5 launch is not just another chatbot launch. Seems more like Cursor is getting its own dedicated model pipeline.
That's important because Cursor probably has the richest coding interaction data in the marketplace. This is not static code. It is actual developer sessions. What developers ask, which diffs they accept, where agents get stuck, which tool calls work, which changes get thrown away. Cursor says Grok 4.5 was trained jointly with SpaceXAI and used trillions of tokens of Cursor data from active interactions with codebases and software tools.
Then you throw in Colossus. It was estimated to take 24 months to build, according to SpaceXAI. They say they built it in 122 days, then doubled it to 200,000 H100 GPUs. I don't want to turn this into a Musk worship session, because he brings a lot of chaos with him, but one area where he gets the benefit of the doubt is the speed of building infrastructure. If xAI starts to iterate Grok every couple of weeks now, I won't be surprised.
Grok 4.5 is good enough to be very useful. It is not like Fable is dying. On DeepSWE 1.0 it is at 62.0% versus Fable max at 66.1% and GPT-5.5 xhigh at 64.31%. It is right there on Terminal Bench 2.1: 83.3% versus 83.4% and 84.3%.
Is it safe to use on a daily basis? Maybe. Is it cheaper and in Cursor? It is worth trying for sure.
The price is what moves it. Grok 4.5 is $2 per million input tokens and $6 per million output tokens. xAI says it runs at 80 tokens per second and averages 15,954 output tokens on SWE Bench Pro, versus 67,020 for Opus 4.8 max.
Maybe that benchmark advantage gets smaller in real projects. I don't know for sure until I run it myself. But even some of that advantage would add up. Coding agents are not just about solving the absolute hardest task. They are also about how many times you can really afford to let them try.
This is where I think Anthropic is playing with fire.
Claude Code used to feel like a flat rate for coding. That was the trick. You paid and you worked. Now Fable 5 means usage credits or $10/$50 API pricing. Sonnet 5 does not give me the cheap fallback I wanted. Grok does.
What I'm building
Delegate tasks. Get software.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.com