That's pretty surprising. Idk about Grok 4.6, but Grok 4.5 was clearly below Fable, Opus 5 and GPT 5.6 Sol.

logicchains 5 hours ago | parent | next [-]

It's much faster, so if you're not doing something cutting-edge, or you're doing the planning yourself and just using the LLM for implementation, the speed benefit outweighs the extra smarts of Fable/Opus5/Sol.

▲

slowin 2 hours ago | parent | next [-]

I've used all of the models extensively and Grok is only "faster" because it claims to be done minutes after you ask it to do something. It does not produce results anywhere near the Anthropic or OpenAI models, it just hacks a tiny piece of what you ask for and says "I'm done!". I also notice Musk-isms leaking through the model. Multiple times it's told me "this is not a roast" or "I'm not roasting your code". People who use this model because they align with Musk's ideology are doing us all a favor and weeding themselves out of the competition.

▲

rlt 4 hours ago | parent | prev | next [-]

Is there a good way to auto switch between models for planning/implementation?

	▲	phoghed 4 hours ago \| parent \| next [-]
		Pretty much every coding harness allows you to define skills and agents so yes
	▲	rubyn00bie 3 hours ago \| parent \| prev \| next [-]
		In most coding harnesses you can just instruct the agent to do so. When using Fable, I often say things like: > Save your context. Always use a subagent (Opus 5 or GPT 5.6) for performing the implementation and then review the work yourself. You are the orchestrator and coordinator it’s up to you to ensure a cohesive final result. I’ve done more or less the same thing with other agents/versions but with Fable consuming usage credits/tokens so quickly I do it more regularly than usual. I know people who will specify Composer (to my chagrin) as the implementing agent. Addendum: this really goes a long way, and I can use a single chat session for days before I get into the context danger zone and need to compact/summarize.
	▲	esafak 4 hours ago \| parent \| prev [-]
		OpenCode can.

▲

redox99 4 hours ago | parent | prev [-]

Yeah the speed was very nice indeed.

▲

porridgeraisin 3 hours ago | parent | prev [-]

It's a different type of model. In my admittedly judgemental observation, people that aren't the type to configure fully automated harnesses with good tools and skills and verifiers for their infrastructure and are way more interventionist in the way their agent works tend to like grok 4.5 more as the main agent. It's much faster and writes more simple and normal code that aligns a bit more with human written code. As an example, instead of sandboxing and simply verifying output artifacts they manually read and approve edits, suggest different code patterns, and manually approved shell commands. On the flip side it's not as good as fable when you need a relatively complex multi step thing done.

But in my experience, the overall productivity ends up similar, give you are willing to work with it in that way.

The grok build TUI harness is excellent and I really enjoyed using it.

For debugging and such I found it pretty much the same as other models.

Fable's taste in software abstraction and project planning in greenfield setups[1] is unmatched in my experience. Sol is OK. My primary use is launching tens of experiments that have to smartly use a limited pool of GPUs.

I use fable to start off the experiments, decide checkpoints, gpu alloc, where to sacrifice precision for performance, and then grok4.5 to iterate, tune, debug, eval, etc, within the abstraction and setup that fable initiated. I have fable write simple scripts that are then wrapped in skills for grok to use. Speed for that loop is extremely important for me, since I also apply human judgement there and I don't like waiting for model output.

I have tried Deepseek and such for the inner agent, but I desperately need multi-modal. Otherwise it's OK, but it tends to use tools less and rambles on and tries to reason with limited information and gets things wrong. Probably a relative la k of tool use posttraining. Gemini flash limits in google ai pro are too low for me to use to compare.

I use anthropic and openais models through grants and so can't compare subscription plan token budgets, but supergrok's budgets are satisfactory.

[1] aside, I have not yet met a model that continues off of a human codebase and actually follows the patterns reliably long term. Eventually it's all slop.

	▲	redox99 2 hours ago \| parent [-]
		Yeah [1] is really a thing and SlopCodeBench (https://www.scbench.ai/) kind of measures that. You need to manually push models to clean up the slop every now and then otherwise it becomes chaotic. And every change with LLMs is always extra lines.