I'm going to say something that will annoy both camps: neither open-source nor proprietary AI models are universally better. We run both in production every day, and the choice is almost always boring and practical.

Here's what actually matters when you're picking a model for a real business application.

The cost question everyone gets wrong

People compare API costs per token and think they've done the math. They haven't. A GPT-5 API call might cost 3x more per token than running Llama 4 on your own infrastructure. But "your own infrastructure" means GPU servers, DevOps time, monitoring, security patches, and that one engineer who has to wake up at 3 AM when the model server runs out of memory.

We did the math for a mid-size insurance company last year. They were processing about 5,000 documents per day. The API route (GPT-5) came out to roughly €4,200/month. Running Llama 4 on two A100 GPUs? About €3,800/month in cloud compute — plus roughly €2,000/month in engineering time for maintenance, monitoring, and the occasional fire drill.

Llama 4 was technically cheaper per inference. GPT-5 was cheaper in total cost of ownership.

But — and this is important — that calculation flips at around 20,000 documents/day. At higher volumes, the fixed cost of infrastructure gets amortized and open-source wins on pure cost.

Where proprietary models still win

1. Zero-setup time to production

You get an API key and you're live. No GPU provisioning, no model serving infrastructure, no CUDA driver compatibility nightmares. For a proof of concept or a low-volume application, this is enormous.

2. Reasoning tasks with high stakes

For complex multi-step reasoning — contract analysis, regulatory compliance checks, financial modeling — GPT-5 and Claude Opus 4 still outperform open-source alternatives. The gap has narrowed significantly (Llama 4 Maverick is genuinely impressive), but it's still there for the hardest tasks.

3. Multimodal without the pain

Need to process images, PDFs, audio, and text in the same pipeline? Proprietary APIs handle this seamlessly. Getting Llama 4 to do the same thing requires stitching together multiple models and dealing with format conversion headaches.

Where open-source wins clearly

1. Data privacy requirements

If your client's contract says "no data leaves our infrastructure," that's the end of the conversation. Llama 4, Qwen 3.5, Mistral 3 — they all run on your own servers. No data goes anywhere. The enterprise AI platforms we build can run the same way, on your own infrastructure.

We have three financial services clients where this was the only deciding factor. The model quality was secondary to the compliance requirement.

2. Fine-tuning for your domain

You can fine-tune Llama 4 on your company's specific data. You can't fine-tune GPT-5 (not meaningfully, anyway — there's a fine-tuning API but it's limited). If your use case requires deep domain knowledge — medical terminology, legal jargon, manufacturing-specific codes — fine-tuning is usually the path to best performance.

We built a document processing system for a large logistics company using a fine-tuned Llama model. After training on 8,000 of their actual invoices, it outperformed GPT-5 on their specific document formats by about 12% accuracy. GPT-5 is smarter in general. Their fine-tuned Llama is smarter about their invoices.

3. Predictable costs at scale

Once you've set up the infrastructure, your costs are fixed regardless of usage. This matters enormously for high-volume applications. One of our clients processes 50,000 customer support tickets per month. On a per-ticket basis, their self-hosted Qwen 3.5 instance costs about one-fifth of what the equivalent API calls would.

What we actually recommend

Here's our default decision framework:

  • Under 1,000 requests/day, no sensitive data: Use proprietary APIs (GPT-5 or Claude Opus 4). The engineering savings outweigh the per-call cost.
  • Under 1,000 requests/day, sensitive data: Use a hosted open-source model on a private cloud instance. Higher setup cost, but the compliance benefit is worth it.
  • Over 5,000 requests/day: Run the numbers. At this volume, open-source on dedicated infrastructure usually wins on total cost.
  • Need domain-specific accuracy: Fine-tune an open-source model. No proprietary API can match a model trained on your actual data.
  • Need fastest time to market: API-first, always. You can migrate to self-hosted later if the economics justify it.

The hybrid approach

Most of our production systems use both. Here's a real example:

A customer support system we built routes incoming tickets through a triage step. Simple categorization and routing? That's handled by a self-hosted Qwen 3.5 instance (fast, cheap, runs 24/7). Complex tickets that need nuanced response generation? Those get escalated to Claude Opus 4 via API.

The result: 80% of tickets are handled by the cheap model. 20% use the expensive one. Total cost is about 40% of what it would be if everything went through the API.

That's not ideology. That's engineering.