Open vs proprietary AI models: which one should your SMB choose?

The choice is no longer binary — and that is good news

Until recently, the reasoning was simple: open models were toys, proprietary models were the only ones usable in production. I started like everyone else: GPT for everything, rising bills, then discovering open models, then moving to a hybrid approach.

What has changed this year is significant enough that SMB leaders should rethink their strategy.

OpenAI launched GPT-5.6 on July 9 — three tiers: Sol (flagship, beats Claude Fable 5 on Agents' Last Exam benchmarks), Terra (balanced general-purpose model), and Luna (economical). The performance-to-price ratio has improved, but the dependency on US API infrastructure remains unchanged.

Anthropic answered with Claude Fable 5, matching GPT-5.6 Sol on complex reasoning and code generation. Both are excellent — and both raise the exact same data sovereignty questions.

Meanwhile, open models have caught up. Mistral has released open-weight models (Mixtral 8x22B, Mistral Nemo, Small) under Apache 2.0. Meta's Llama 3 delivers performance rivalling GPT-5.6 Terra on most common tasks. DeepSeek, Qwen, Gemma — the list grows every month. I now host models for clients on tasks I would have thought impossible without an API two years ago.

Bonsai 27B, announced in mid-July 2026, pushes the logic further: 27 billion parameters designed for modest hardware. If this model delivers on its promises, the question "do I need a GPU?" disappears for most SMB use cases. A €10/month VPS becomes sufficient.

But the reverse is also true. Proprietary is not dead. GPT-5.6 Sol and Claude Fable 5 keep improving. For certain tasks — complex reasoning, contract analysis, creative generation — the gap with open source still exists. The difference is you can now choose precisely when to accept it.

This is not an ideological debate between open source and proprietary. It is an engineering and economics choice. And like all engineering choices, it depends on your context.

What I have learned deploying models for my clients

I have guided about a dozen SMBs and professional firms through choosing and deploying their AI infrastructure. Here is what comes up consistently.

First lesson: data sovereignty is the real filter, not cost.

Before comparing prices, ask this question: can my data leave the country?

The US CLOUD Act allows US authorities to compel any US-incorporated company — OpenAI, Anthropic, Google, Microsoft, Amazon — to hand over data stored on their servers, even if those servers are in Europe. GDPR imposes safeguards, but the CLOUD Act can override them.

I had a client in the insurance industry who was processing claims data — names, addresses, amounts, medical history — through GPT-4. Nobody had realised this information was leaving for the United States. We switched to Mistral Nemo hosted on Scaleway in France. The monthly cost dropped by a factor of 15 and the legal problem disappeared. The model is less performant on creative tasks, but for claims analysis — classification, extraction, summarisation — it does exactly the same job.

Second lesson: 80 % of SMB tasks do not need a frontier model.

When I deploy a Hermes agent or an n8n workflow for a client, the tasks are rarely exotic: classify documents, extract data, generate standard replies, analyse sentiment, summarise a conversation.

I have tested this systematically: on these tasks, a self-hosted 7 to 12-billion parameter model (Mistral Nemo, Llama 3.1 8B) delivers results comparable to GPT-5.6 Terra in 80 to 90 % of cases. Not "almost as good." Just as good. The difference is not measurable in daily use.

Third lesson: proprietary API costs follow an exponential curve that leaders underestimate.

One client started with GPT-5.6 Luna for a customer support agent. €30 per month at first. Then €80. Then €200. Six months later, they were paying €450 per month without anyone having approved the increase. The model had not changed. The volume had.

I installed Mistral Nemo on a VPS at their office. Cost: €9 per month. The local model handles 90 % of requests. The remaining 10 % — complex cases where the customer is upset and the response needs nuance — are routed to GPT-5.6 Luna. Monthly API bill: €25. Total: €34/month instead of €450. And sensitive data (contact details, order history) no longer leaves the server.

This is the kind of decision that does not show up in a benchmark.

What remains proprietary territory

It would be dishonest to claim open models replace everything. Some cases remain better served by proprietary APIs — and I use them myself.

Complex, multi-step reasoning. Analysing a contract that requires cross-referencing multiple clauses, checking regulatory consistency, and producing a synthesised argument — GPT-5.6 Sol and Claude Fable 5 are significantly better.

Advanced code generation. For writing complex functions or debugging obscure issues, Claude Code and GPT-5.6 are worth their cost. I use them daily.

Minimal latency. A proprietary API responds in 200-500 ms. A self-hosted CPU model — without GPU — responds in 1-5 seconds. For a real-time customer chatbot, the difference matters.

But these cases represent a fraction of total volume. The key is not to pay premium prices on the 80 % that do not require it.

The hybrid strategy: how I actually implement it

Here is the architecture I install for my clients today. It is battle-tested enough to fit in a few paragraphs.

The local layer (volume): A CPU server at €6-10/month on Scaleway, Hetzner or OVH. A Mistral Nemo 12B or Llama 3 8B model in 4-bit quantized format. Ollama or vLLM to serve the model. The server runs 24/7 and handles background tasks — classification, extraction, summarisation.

The API layer (escalation): An OpenAI or Mistral API account for complex requests. No fixed subscription — pay as you go. Volume is low enough that the bill stays under €30/month.

The routing layer (the brain): In n8n or the Hermes agent, each request arrives with a confidence score. If the local model responds with confidence above 85 %, the reply goes directly. If confidence is between 60 and 85 %, the response is proposed to a human for quick review. Below 60 %, the request is routed to the proprietary API.

This is not theory. I have this system running at three clients currently, with variations. The oldest has 14 months of production uptime without incident.

A concrete example

Consider an accounting firm I work with. 8 employees, 200 SMB clients. Their need: automatically classify accounting documents that clients send by email — invoices, expense reports, bank statements, contracts.

Before: an assistant spent 15 hours per week manually sorting email attachments. Sometimes 250 to 300 documents per week.

What we set up:

• An OVH server at €7/month, Mistral Nemo 12B
• An n8n workflow that fetches emails, extracts attachments, sends them to the model for classification (supplier invoice / expense report / bank statement / contract / other), then files each document in the correct Drive folder
• Documents the model classifies with confidence below 70 % (about 8 % of volume) go into a review queue the assistant handles in 5 minutes

Result after 6 months:

• The assistant spends 2 hours per week instead of 15
• The classification error rate dropped from ~5 % (manual, fatigue) to < 1 %
• The local model handles 92 % of documents without human intervention
• Infrastructure cost: €7/month server + €2/month Mistral API for complex cases

If we had routed everything through GPT-5.6 Luna, the monthly bill would have been around €150 for the same volume — and client data (financial statements, payslips, bank records) would have gone to the United States via Azure. The firm was not willing to take that risk, and rightly so.

5 questions to decide

If you are reading this wondering where to start, here are the questions I ask my clients during the first meeting.

1. Can my data leave the country?

If the answer is no (HR, medical, banking, customer PII), the question is settled: self-hosted or Mistral via API hosted in France. GPT-5.6 and Claude Fable 5 excluded.

2. How many AI requests per month?

Fewer than 10,000: a proprietary API may be simpler. More than 50,000: self-hosting becomes essential. Between the two, run the numbers.

3. What latency is acceptable?

Real-time chatbot: proprietary API. Batch processing: self-hosted.

4. What level of performance do you actually need?

Not what benchmarks show — what your daily operations require. For classifying invoices, a self-hosted 8B model is as good as GPT-5.6 Terra. For generating creative sales proposals, proprietary remains superior.

5. Do you have someone to maintain the model?

A self-hosted model needs updates, monitoring, and troubleshooting. If no one on your team has those skills, a proprietary API or a specialised partner is the right path.

The choice is strategic, not technical

I end my audits with a sentence clients remember: "The right AI model for your business is not the one with the best benchmark. It is the one that keeps your data in Europe, will not bankrupt you in six months, and gets the job done for 90 % of your use cases."

Open and proprietary models are not in competition. They are complementary. The question is not which one to choose, but what to send where — and having the infrastructure to do it.

This is exactly the kind of decision people hire me for. Not to install a server — to decide what goes on it, what goes elsewhere, and what stays under human control. And to make it all fit in an SMB budget.