The title asserts there is minimal downside to switching to open models, but the article provides zero evidence that this is true, and the author hasn't even attempted it yet. The end of the article states "I’m hoping it’s going to be minimal".

I wonder if I can get a post to the front page with the title: "There are no real barriers to humans colonizing Mars next month". And at the end, "I'm hoping there are no real challenges."

I know LLMs move at the speed of light (especially these past few quarters), but if Opus and GPT "a few months ago" were really like open weight models, then there's really no reason to not switch, especially for those who were using these models a few months ago.

Your codebase didn't change, so use the open weight model. Don't move the goalposts.

How did we get from prising software freedoms to this?

[1] It seems inevitable that decent local models will be possible as the technology and the hardware is improving at a rate beyond the growth of the knowledge base to be distilled.

> I’m hoping it’s going to be minimal.

I have multiple subscriptions and I pay per token to try out different LLM providers through OpenRouter. I also run open weight models locally.

I just can’t agree yet. The models from Anthropic and OpenAI really are that much better than anything else. The open weight models must be universally benchmaxxed across the board because my real world experience with them is very different than what the benchmarks imply. I get downvoted a lot for speaking about my experience because I don’t think it’s the reality that people want to hear right now, but it’s true for complex work.

I do think there are a lot of easier tasks that can be handled appropriately by the open weight models in the hands of a skilled operator. If an entire job is simple enough that you wouldn’t hesitate to hand it off to a junior with a little supervision then any model will do. However for a lot of the work I do, even Opus 4.8 on Max requires a lot of attention and extra steering and review to keep it on track. Fable did, too, though to a lesser degree. When I try to use the big open weight models (hosted, because they’re not running at reasonable speeds locally at a quantization I can tolerate) it feels like I spend more time waiting while they burn tokens for output that I probably have to reject anyway, at least for the bigger tasks. I wish they were there, but that’s not the case yet.

Having played a bit with Fable, reinforced the above.

I think it would be pretty neat to launch a service helping people who wanted to participate in something like that locate one another.

1. Unfortunatly in my tests the open models do not (yet?) rival, at least Claude Opus, for software development/engineering and adjacent tasks.

2. Enjoy while it lasts. I'll be genuinly amazed these open models will not be declared 'illegal' under some security pretense by the end of the year. And I say 'pretense' because the primary driver will be regulatory capture and industry protectionism.

Don’t get me wrong. I wish I could run a local model and be happy about it. At the moment, I’m not.

Sure, there may be some cases and reasons for local models and industry is so large they will continue to make progress and gather economic value and users for specific use case; but frontier will command vast majority of the economic value distinct from Linux and open source where the model created better than proriatary economic incentives around development

Personally I haven't seen any productivity gain since Opus 4.5 times.

But: I can't fully get behind the opinion that (so called) "open source models" are simply superior and will be in the future, because when I asked some models who they are, they answered with "I am Claude from Anthropic", which could mean they have been trained by exfiltrating Claude.

I have NO moral objection to this, as Anthropic and "Open""AI".also trained their models on anything they could get their hands on.

It's more about the question: can and will these models be updated, even if Anthropic et al fail. Who's gonna pay for training then? What's their incentive? Have we reached a plateau?

Article: "I’m hoping it’s going to be minimal"

So while it is not complicated and certainly something that can be solved, it is not plug and play.

That being said, we switch to open weight models earlier this month and the results has been more than positive so far. The cost savings are also hard to dismiss.

For a while during this era, I used to port my laptops windows installation into a virtual machine that can run on Linux. It took a bit of hacking away but I could usually do it in a day or two. Then its all Linux with the windows vm being used for the microsoft stuff.

I do have to admit I have recently begun wishing I could pay five dollars a month for a "just answer the fucking question" plan that would give me results without the guardrails and without the constant simpering and ego-stroking. I keep finding myself going a quick evaluation of "is it faster for me to skim search results myself or to construct an elaborate narrative to make an AI give me a real answer".

1. Evals that can quickly tell you how much downside there is to switching 2. Something like OpenRouter that can help you run those evals quickly

Now #2 is starting to become popular, and I think we'll soon see more people adopting a model-agnostic approach. Of course, there will still be high-intelligence use cases where nothing comes close to Claude or GPT.

Think about it, if you want to use an open source model to run an agent over long horizons, you need someone to offer it reliably. And if all compute goes to frontier models there is very little left for even startups to build agentic companies on top of these open source models.

Whatever reason people have to run those (cheaper? backwards compatibility once you get something running) surely applies to the open models too, maybe even more so.

A $10,000 RTX 6000 Blackwell card will pay for 500 months of Claude or Codex, which is 40 years worth of compute. Obviously they are going to raise their prices, my prediction being to $200-500/month, but that still makes them at least years of compute and they scale very well with more traffic. Single GPUs do not, they are pegged at 100% and good luck getting it to answer multiple queries at the same time.

and what hardware are you using?

I like the Linux analogy, I struggled with Linux way back.

But it doesn't have to be an "AI company". It's just a compute service. The companies that offer web hosting could get into this.

I enjoyed the first part though

There are several benefits:

- we cut AI spending by thousands

- there is one AI server and starting different sessions for each user, one memory/skills/etc and everybody is involved into reviewing what went wrong and why. Harness finally makes sense and pays off more.

- we can trust that the models are those that we run and not black boxes

- no more money flowing to US narcissistic entrepeneurs and no more business being tied to US legislation

Not gonna lie, GPT 5.5 Pro and Fable 5 were a tiny bit ahead, especially on longer vibecode-style tasks, but it's just not worth it.

That's why I'm using eurouter.ai with the following routing rule for all my requests:

  {
    "model": "glm-5.2",
    "models": [
      "deepseek-v4-pro",
      "deepseek-v4-flash"
    ],
    "provider": {
      "allow_fallbacks": true,
      "data_collection": "deny",
      "data_residency": "EU",
      "max_retention_days": 0,
      "eu_owned": true
    }
  }

Personally, I find it morally unacceptable to use U.S. AI tools, because I do not want to support them financially and thus support the crimes they are involved in[1].

[1]: https://news.ycombinator.com/item?id=48512339