Six months ago, if you asked most enterprise CTOs whether they’d trust an open-source AI model for production workloads, they’d have laughed. Not anymore.

The speed at which open-source large language models have closed the gap with proprietary offerings from OpenAI, Anthropic, and Google is genuinely stunning. And I don’t think the big labs saw it coming this fast either.

The benchmarks tell the story

Let’s talk specifics, because vague claims about “catching up” don’t help anyone.

Meta’s Llama 3.3 70B now matches or exceeds GPT-4’s original performance on most standard benchmarks — MMLU, HumanEval, GSM8K. That’s a model you can download and run on your own hardware. Mistral’s Large model is competing head-to-head with Claude 3.5 Sonnet on coding tasks. And then there’s DeepSeek — the Chinese lab that dropped DeepSeek-V3 and basically forced the entire industry to recalibrate what’s possible at a given compute budget.

DeepSeek-V3’s mixture-of-experts architecture hit GPT-4-class performance while reportedly being trained for a fraction of the cost. Their R1 reasoning model followed shortly after, and it’s competitive with OpenAI’s o1 on mathematical reasoning. These aren’t minor achievements. They represent a fundamental shift in who can build state-of-the-art AI.

The latest round of open-weight models from Alibaba’s Qwen team and 01.AI’s Yi series have pushed things further still. On MIT Technology Review’s analysis of the AI landscape, even the mid-size open models (7B-13B parameter range) are now handling tasks that required 70B+ parameter proprietary models just eighteen months ago.

Why this matters for enterprise

Here’s where it gets interesting for anyone making real purchasing decisions.

If you’re an enterprise running GPT-4 or Claude through an API, you’re paying per token. For heavy workloads — document processing, customer service, code generation — those costs add up fast. A mid-size company doing serious AI integration might be spending $15,000 to $50,000 a month on API calls alone.

Now imagine running an equivalent open-source model on your own infrastructure. The upfront investment is higher — you need GPUs, you need ML ops expertise, you need to handle fine-tuning yourself. But the marginal cost per query drops dramatically. For high-volume use cases, we’re talking about 60-80% cost reductions once the infrastructure is amortised.

That’s not a rounding error. That’s the kind of number that changes procurement decisions.

The real competitive advantage has shifted

I think we’re watching a fundamental restructuring of where value sits in the AI stack. It used to be that the model itself was the moat. You paid OpenAI because nobody else had anything close. That’s just not true anymore.

The competitive advantage is moving to three places:

Fine-tuning and specialisation. A general-purpose open model fine-tuned on your specific industry data can outperform a larger proprietary model that doesn’t know your domain. I’ve seen this firsthand with legal document review and medical coding tasks.

Infrastructure and tooling. The companies that make it easy to deploy, monitor, and iterate on open models — think Hugging Face, Anyscale, Together AI — are capturing enormous value.

Data quality. As always, the companies with the best proprietary data win, regardless of which base model they’re running. Open-source models just made the base model commodity. Your data is what differentiates you.

What the big labs are doing about it

OpenAI’s response has been to push harder on reasoning (the o-series models) and multimodal capabilities. Anthropic is leaning into safety and reliability. Google’s betting that tight integration with their cloud ecosystem keeps enterprises locked in.

These are all reasonable strategies. But they’re defensive strategies. The frontier is still moving, and each new open-source release compresses the window where proprietary models hold a clear advantage.

I think the most telling signal is pricing. OpenAI has dropped prices aggressively over the past year — multiple times. That’s not generosity. That’s competition.

My take

Here’s what I actually think is happening: we’re heading toward a world where the base model is essentially a commodity, similar to databases or operating systems. Yes, there’ll be premium offerings. Yes, some will be better than others at specific tasks. But the gap between “good enough” and “best available” will be small enough that most organisations won’t pay a 5x premium for marginal improvements.

That’s good news for basically everyone except the companies whose entire business model depends on that gap staying wide.

For enterprises, it means the “wait and see” approach to AI adoption is getting harder to justify. The cost of experimentation has dropped through the floor. You can spin up a Llama 3.3 instance, fine-tune it on your data, and have a production-ready system for a fraction of what this would have cost two years ago.

The open-source AI movement isn’t just catching up. It’s rewriting the rules of the market. And if you’re still treating AI adoption as a question of “which vendor do we pick,” you might be asking the wrong question entirely.

The better question is: what can we build now that the models are essentially free?