LIVE
US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|
Intermediate⚖️

Open source vs proprietary: which LLM to choose?

Mistral, Llama, DeepSeek vs GPT, Claude, Gemini: open or closed? We help you decide based on your volume, GDPR constraints and real budget (with TCO calculation).

12 min readPublished May 5, 2026· Updated September 17, 2026

In one sentence

Open-source LLMs (Mistral, Llama, DeepSeek) and proprietary LLMs (GPT, Claude, Gemini) each have their strengths. The right choice depends on 3 criteria: control, total cost, and raw performance. We'll help you decide.

⚖️
The analogy that works
It's like choosing between buying a house (open source) or renting a premium flat (proprietary). The house gives you total control, but you have to manage everything. The flat, everything's included, but you depend on the landlord. Neither is "better", it depends on your life.

🚀 Want to compare side-by-side?

30 open-source AND proprietary models compared: price, performance, language, licences.

View comparison

Open source: what exactly is it?

An open-source LLM is a model whose weights (the billions of parameters) are publicly downloadable. You can:

  • Run it on your server, your computer, or even in the cloud
  • Modify it (fine-tuning on your data)
  • Redistribute it (according to the licence)
  • Inspect it (total transparency)
Be careful with the term 'open source'
Not all "open" models are truly open source in the strict sense (OSI-approved licence). Many use custom licences with restrictions: - Llama 4: permissive licence but prohibits use by very large companies (>700M MAU) - Mistral: Apache 2.0 (truly free) for "research" models, commercial licences for others - DeepSeek: MIT (totally free) Always check the licence before using in production.

Proprietary: what exactly is it?

A proprietary LLM is a model accessible only via an API (or chatbot) hosted by the provider. You never have access to the weights. The provider maintains total control over:

  • The infrastructure (latency, availability)
  • Updates (can change behaviour without warning)
  • The price (can increase)
  • The data (yours necessarily passes through it)

The big match: Open vs Proprietary

Open source vs Proprietary

 🔓Open source🔒Proprietary
Data controlTotal (your servers)Limited (their servers)
Raw performance (top level)Very good (Llama 4, DeepSeek)State-of-the-art (GPT-5, Opus 4.7)
Cost (low volume)High (infra to pay)Very low (pay-per-use)
Cost (high volume)Very low (~£0.01/1k tokens)Expensive (~£0.15/1k tokens top API)
Time-to-marketSlow (setup, monitoring)Immediate (1 line of code)
Customisation (fine-tuning)Total and freeLimited or expensive
GDPR complianceEasy (EU/local)Complicated (US transfers)
UpdatesYou alone decideImposed by provider
Support / DocumentationCommunity (variable)Professional (SLA, team)

Cost: the great illusion

Many think open source = free. That's false. Here's the reality:

Monthly cost for 100M tokens/month

Llama 4 self-host (1 H100 GPU)2 500€/month
Claude Opus 4.7 API1 800€/month
GPT-5 API1 500€/month
Mistral Large API600€/month
Llama 4 self-host (1 shared GPU)400€/month

Reading the chart:

  • At low volume, proprietary APIs are cheaper
  • At high volume (100M+ tokens/month), open source becomes competitive
  • The break-even is around 50M tokens/month (~5,000-10,000 active users)
The true cost of open source
Running an open-source LLM isn't just downloading a file. You need: - GPU (graphics card): £1,500/month for an H100 on AWS, £800/month collocated - ML engineer: monitoring, optimisation, updates (1 day/week = £1,500/month) - Security: firewall, access, audit - Storage / network: ~£100-300/month → Minimum total: ~£3,000-5,000/month for a serious deployment. Below that, proprietary API is more cost-effective.

Practical cases: what to choose?

📚6 common situations and the right decision

1. "I'm testing a POC in 1 week"

Choice: Proprietary API (Claude, GPT, Mistral) Why: you want immediate results, no infrastructure to set up. The cost for 1 week of tests is negligible.

2. "I have a startup with 500 active users"

Choice: Proprietary API Why: your volume remains low (<10M tokens/month). Open source would cost more in infrastructure than an API.

3. "I have 50,000 active users in production"

Choice: Self-hosted open source (or hybrid) Why: your volume justifies the infrastructure investment. Savings of 50-70% vs proprietary API at this stage.

4. "I process medical / financial data"

Choice: Local open source OR proprietary Enterprise offering with DPA Why: GDPR. Your data must not leave a controlled environment.

5. "I want to customise the model (style, industry vocabulary)"

Choice: Open source (fine-tuning on Mistral or Llama) Why: fine-tuning on proprietary API exists but is limited, expensive, and depends on the provider.

6. "I want the best possible performance"

Choice: Proprietary (Claude Opus 4.7 or GPT-5) Why: proprietary "frontier" models maintain a 6-12 month lead over the best open source.

The best open-source LLMs in 2026

Top open source 2026

 🏆Model🎯For whom?
Mistral Large 3 (Mistral AI, FR)123B params - excellent FR/ENGDPR-friendly Europeans
Llama 4 (Meta)70B / 405B params - generalistClassic US usage
DeepSeek V3671B MoE (37B active) - top performanceResearch, reasoning
Qwen 3 (Alibaba)Multilingual, fastAsia, multilingual
Gemma 3 (Google)Small model (2B-9B)Edge, mobile, low-resource

The evolution open vs proprietary

The gap narrows every year

  1. GPT-3 vs GPT-J

    Huge gap. GPT-3 (proprietary) crushes GPT-J (open). 18 months behind.

  2. GPT-4 vs Llama 2

    Gap reduces to ~12 months. Llama 2 70B is competitive with GPT-3.5.

  3. Claude 3.5 vs Llama 3

    Llama 3 70B reaches GPT-4 level. ~6 months gap.

  4. DeepSeek V3 surprises everyone

    Chinese open model at 671B MoE, GPT-4o level for ~6% of training cost.

  5. Open source catches up

    Llama 4, DeepSeek V3, Mistral Large 3 are at 90-95% of frontier proprietary models. The gap is now 3-6 months.

The hybrid strategy: best of both worlds

🔀
The professional approach 2026
Most serious companies adopt a hybrid strategy: - Open-source model for 80% of queries (FAQ, classification, short summaries) - Proprietary API for 20% of complex queries (reasoning, expert code) Result: 60-70% savings vs full proprietary, with equivalent performance on critical cases. Tool for this: an LLM router (LangChain, LlamaIndex) that automatically dispatches.

Pitfalls to avoid

3 classic mistakes
1. Underestimating the cost of open source Many migrate to self-hosted "to save money" and end up paying more than the API. Always calculate TCO (Total Cost of Ownership) over 12 months. 2. Choosing open source for "freedom" without needing it If you don't have GDPR constraints / fine-tuning / gigantic volume, proprietary API is almost always more efficient. 3. Choosing proprietary without calculating lock-in risk If OpenAI doubled its prices tomorrow, would your product still be viable? For strategic uses, keeping an alternative open source ready is wise.

The metaphor that sums it all up

🍔
McDonald's vs home cooking
Choosing between open and proprietary is like choosing between eating at McDonald's or cooking at home. McDonald's (proprietary): it's fast, predictable, you pay for each meal. If you eat there once a week, it's cost-effective. If you eat there 3 times a day, it becomes expensive and not great for health (vendor lock-in). Home cooking (open source): it's more work (shopping, cooking, dishes), but long-term it's cheaper, customisable, and you know what's in it. Ideal if you eat often and know how to cook. Many professionals do both: batch-cook open source for daily use, proprietary restaurant for special occasions.

Key takeaways

  • Open source ≠ free: count ~£3,000/month minimum for a serious deployment
  • Proprietary is the right default choice to start / low volume
  • Open source becomes cost-effective above 50M tokens/month
  • Hybrid strategy = optimum for most cases
  • Always keep an alternative to avoid vendor lock-in

The right choice depends on YOUR context: volume, criticality, GDPR constraints, technical team. No universal answer.

🧠 Quiz
Question 1 of 3

From what monthly volume does open source generally become cost-effective?

Going further

Tags
Open SourceAPIComparatifMistralLlama

Read next