Dieser Blogpost ist auch auf Deutsch verfügbar
- Rising AI prices are annoying, but dependence on a small number of U.S. providers is the bigger risk.
- Open-weight models and EU-hosted services offer stronger privacy guarantees and reduce vendor lock-in.
- Some open-weight models are already good enough for day-to-day software development.
- We don’t need to abandon Claude or GPT, but we should actively explore alternatives before we need them.
Anthropic, OpenAI and Google have been at the front of the AI game for a long time and they’ve also managed to convince the software development industry how useful agentic software engineering can be. Agents have particularly disrupted the IDE industry. A year ago, I would have fought tooth and claw if someone had tried to take away my IntelliJ license. Now I increasingly use my IDE only to review the code which has already been generated and would just find another tool if the JetBrains licenses go through the roof.
Rising Costs
But now that Anthropic and OpenAI have likely found product market fit, they have also begun raising the pricing. The GPT models used to be quite a bit cheaper to use than the Anthropic models, but with the release of GPT 5.5, it became the most expensive model on the market (at least until Anthropic released its new Fable model).
Most dissatisfying to me was not that the prices were rising. We are dealing with U.S. companies who are out to make a profit, so I fully expected that.
What dissatisfied me was the feeling that I was paying more money without getting dramatically better results in return.
I think it has to do with how you use the models.
If you have a huge code base with lots of different code styles and you want your agent to be productive, you may need the latest huge cutting edge model.
If you are having your agent generate a lot of code or orchestrate a lot of tasks for you, you may benefit from the latest huge cutting edge model. I personally consider this approach to software development to be a bit experimental and have decided to let others approach that in their code bases and see how they fare before I implement it in my own.
I still work most effectively when I break the scope of my work down into chunks that are small enough that I am able to review them.
And for chunks of work with that size of scope?
I haven’t seen an improvement in model quality since GPT 5.3 or Sonnet 4.6.
Yet I am still expected to shell out more money to the same providers, even when they remove those older, cheaper models that are fully sufficient for my use case from their subscriptions.
Digital Sovereignty
Cost is the most urgent concern, but it is not the most critical concern. The U.S. government seems intent on controlling the models that other nations have access to. In a sense, I appreciate that the U.S. government is interfering so overtly because it reminds us that we are increasingly dependent on technology which could be gone tomorrow.
Anthropic and OpenAI have, of course, a vested interest in ensuring that we stick with their products. But this is also not without concern. OpenAI has made decisions that are questionable on an ethical level (like their decision to work together with the U.S. military). Anthropic has positioned itself as the good guy in the AI field – but I can’t help thinking of the old Google motto “Don’t be evil” and since Anthropic is still a for-profit Silicon Valley company, I’m skeptical that they will care about others' interests when they conflict with their own.
There is also a huge conflict of interest when it comes to data processing agreements with companies that train AI models because the session data we create while working with agents is exactly what those companies need to make their models even better. We just have to trust that they are going to follow the rules (an idea which makes me slightly nervous). In general, I think for normal coding projects the risk of data misuse is likely low because I’m not dealing with sensitive customer data or credentials, but I would still prefer if those companies didn’t have access to my data to begin with.
Exploring Alternatives
So what can we do?
U.S. Models Hosted in Europe
If you want to use Anthropic models, you can look at accessing them via Amazon Bedrock. If you want to use OpenAI models, it’s also possible to access them through Azure OpenAI.
Both of these options are better from a data privacy perspective, because the data is not sent directly to OpenAI or Anthropic and there are options for hosting the models within the EU. You are, however, still dependent on large U.S. corporations, so there is still a risk of government interference.
The large U.S. models are the most well-known, but that does not mean that they are the only models that we can use.
Local Models
One option is to use local models, which are optimal from a data privacy perspective because the data never leaves our local machine. If we are privileged enough to already own expensive hardware with a lot of RAM, we can run the models ourselves with no extra cost apart from the added electricity when our computer is running on all cylinders.
I’ve tried out the Qwen3.6–35b-a3b model running on my 64GB MacBook Pro for agentic coding. It is an interesting model because it attempts to optimize the model inference so that a relatively large model (35b stands for 35 billion parameters) can be run on a smaller architecture (a3b means that only 3 billion parameters are loaded at any given time).
For day-to-day coding tasks, local models don’t quite work for me because the context window is still limited and the scope of the tasks that the model can handle is smaller than what I am used to. This isn’t strictly an issue with the local models themselves. I know colleagues who have coded quite successfully with local models and some people have used local models to do agentic programming in airplanes or trains where the internet is severely limited.
The issue for me is that local models are not (yet) a drop-in replacement for a larger Claude or Codex, so using them would require me to change my day-to-day workflow which is currently too much friction when I have other options available.
Open Weights Models Hosted in the EU
Another option is to look at other AI providers which provide hosting for different models in the EU. This allows us to have access to models which are more optimal from a data privacy perspective and we can avoid vendor lock-in.
I’m very grateful for my INNOQ colleagues who have set up an internal AI platform for us to test out many different models we can use as an alternative to OpenAI and Anthropic. We only have to configure credentials for one provider in an agent harness like opencode. This currently gives us access to 22 different models (and it feels like we receive access to a new one every few weeks). We currently have access to Mistral models as well as open weight models hosted on StackIT, Tensorix, Scaleway, and Novita. Most of the models are hosted in the EU, and we also have a data processing agreement which allows us to use the models as a drop-in replacement for Claude or GPT (we only use models for customer projects when we have a data processing agreement and explicit consent from the customer to use AI in the project).
Today, I tried out the GLM 5.2 open weights model published by z.ai on our internal platform and I am very impressed with the results. The model follows my instructions, produces code with a similar style to the rest of the code base. It even followed our repository convention of using German for the test names and descriptions, which is something I’m continually having to argue with Claude and Codex about. It’s the first time I’ve tried a model and had the almost immediate conviction that it could be a drop-in replacement for my day-to-day work (sighs in relief).
Because models like the GLM and Qwen models are developed by Chinese companies, there are still some concerns about digital sovereignty because the Chinese government could interfere and prevent us from getting access to newer models from the Chinese AI companies. But these models are open weights, which means that the model parameters are available publicly, even though you don’t know all the data that the model was trained on. This gives us more security from a digital sovereignty perspective, because if we find an open weights model which works for us and is hosted in the EU, a state actor can no longer interfere with our access to that model for our day-to-day work.
The Next Steps
We should take a step back and see if there are steps that we can take to gain a bit of independence from the U.S. corporate giants.
Instead of using only proprietary agent harnesses like Claude Code or Codex, we can use opencode or Pi.
Instead of having company accounts only for the U.S. providers, we can look at different independent AI providers like Tensorix who can provide us with access to open weights models which are hosted in the EU and have a zero retention data policy.
Perhaps you will be surprised – as I was – that other models you have never heard of can actually take the place of the ones everyone is talking about.
Changes to our day-to-day workflow are a bit scary and not without friction. But if we get too comfortable with tooling and models from U.S. providers, we may find the vendor lock-in too difficult to escape by the time we really need to.
If we are already using models from Anthropic, OpenAI or Google, we don’t need to completely throw everything away and start from scratch. But we should keep our eyes open and be curious. The world of AI keeps changing all of the time and we may find many less popular tools which provide us with a better cost/benefit ratio and digital sovereignty at the same time.