The current landscape

Proprietary: OpenAI (GPT family), Anthropic (Claude family), Google (Gemini family proprietary tier). Best raw capability, controlled by their company.

Open weights: Meta (Llama 3), Mistral (Mistral 7B, Mixtral), Google (Gemma), Alibaba (Qwen), DeepSeek. Downloadable weights, runnable locally.

Capability gap

Until 2023, the gap between proprietary and open was huge. In 2024-2026, it narrowed dramatically. Llama 3 405B beats original GPT-4. Mistral Large 2 is competitive with GPT-4 Turbo class. Qwen 2.5 reaches GPT-4o-mini level.

The current frontier still belongs to proprietary (GPT-5.5, Claude Opus 4.7, Gemini 4 Ultra). But for 90% of enterprise use cases, the difference doesn't matter — both serve.

Real cost: not always cheaper open

Common myth: open = free. Reality: running Llama 70B in production costs more than calling GPT-4o-mini API for most use cases. You need to add GPU infrastructure, MLOps, monitoring.

Open is cheaper when: very high volume (millions of daily inferences), have GPU infrastructure already amortized, fine-tuning on specific domain that justifies investment.

The control argument

Beyond cost, open has structural advantages: privacy (data never leaves your control), fine-tuning on private data, not depending on provider (API price changes, deprecations), compliance in regulated industries.

When open makes sense

Regulated industries: banking, healthcare, government. Massive volume: applications with millions of daily inferences. Domain-specific fine-tuning: when you need expertise the base model doesn't have. Compliance: EU/data residency, jurisdictions where US APIs aren't allowed. Edge: mobile apps, IoT, isolated locations.

When proprietary makes sense

Frontier capability: use cases requiring maximum reasoning. Small team: SMBs without MLOps infrastructure. Time to market: validate idea fast without infrastructure investment. Multimodality at top: proprietary leads in voice/vision integration.

Hybrid strategy

Many serious companies adopt hybrid: proprietary frontier (Claude/GPT) for cases requiring max capability, open (Llama/Mistral) for volume and privacy. The architecture supports both: same API gateway, different routes per use case.

Conclusion

The "open vs proprietary" battle isn't zero-sum. Both have a place. The right decision depends on volume, compliance, capability needed and operational maturity. The mature enterprise of 2026 has serious answers to: which use case goes proprietary, which goes open, and why.