Everyone argues over the same question: which model is best? GPT vs Claude vs Gemini, a ranking that changes every month. Meanwhile, a much deeper shift is happening quietly, especially inside large companies: the move to open models.
This isn't about ideology or free-software activism. It's about economics and control. And it's reshaping how a serious company deploys AI.
Closed model, open model: the difference in one minute
A closed model, you rent. It runs on the vendor's infrastructure, you reach it remotely, and you can neither look inside nor modify it. These are the big API models: GPT (OpenAI), Claude (Anthropic), Gemini (Google).
An open model (also called open weights), the vendor publishes. You can download it, run it on your own machines or the cloud of your choice, inspect it, and adapt it to your business. The best known: Llama (Meta), Mistral, DeepSeek, Qwen, Gemma.
The simplest picture: a closed model is a taxi - comfortable, immediate, but the meter is running and the car isn't yours. An open model is buying the car - a cost up front, but it's yours, and you go where you want.
The economics are tipping - but not the way you'd think
The striking number comes from Ollama, the tool through which open models for 85% of the Fortune 500 now flow. Their read: soon 80 to 90% of a company's tokens will run on open models - but they'll account for only 10 to 20% of the budget. In other words, open collapses the cost; frontier closed models stay reserved for the hardest tasks.
But cost is only the front door, not the house. A company moving to open isn't only trying to pay less. It's after control and customization: a model you can host, adapt to your business, govern. Cost starts the adoption; control locks it in. (AT&T has reportedly already shifted 40% of its token consumption to open models.)
The secret: it's not ONE model, it's a team
The winning pattern isn't a single model: it's orchestration. Fast, cheap models - the new generation of "flash models" - handle the bulk of the grunt work; a router only escalates the hardest 10% to the (expensive) frontier model. The analogy speaks for itself: a law firm, where the senior partner delegates to associates and handles only the delicate judgment calls.
The effect on your bill is direct: stop reasoning in cost per token, reason in cost per task. Orchestrated well, the cost of an entire use case can collapse.
Local, cloud, or both
Because they're open, these models can run on your own machines. Document processing, for instance, runs very well locally: cheap, responsive, and your data never leaves your walls. Hard reasoning and code, on the other hand, are still better on the big cloud models.
Recent hardware runs models of 20 to 40 billion parameters without breaking a sweat (sometimes up to ~128) - and some open code models now rival the best closed models on benchmarks. A router arbitrates between local and cloud depending on the task.
The lever is concrete: a "process my documents" use case can see its marginal cost tend toward zero if it runs on machines your company has already bought.
The real brake isn't capability - it's governance
The number-one blocker to open-model adoption in the enterprise isn't "are they good enough?" - for the vast majority of tasks, they are. It's security and governance: where the model comes from, where it runs, who accesses what, is it safe.
The good news is that this problem is methodical, not magical: a selection framework, a deliberate choice between origin and hosting, guardrails. It's exactly the kind of foundation you lay once and reuse across every use case afterward.
What it really changes for you: independence
Here's the underlying argument, and it's strategic. The AI model is becoming a commodity: everyone can rent or download the same one. What stays yours and can't be copied is your context, your data, your ways of working.
Open models turn a principle into a concrete lever. "Don't depend on a single vendor" stops being a slogan: you can host, adapt, and switch models overnight without rebuilding a thing. When everyone's engine puts out a thousand horsepower, the race is no longer won on the engine - it's won on the map. The model is the engine; your map is your company brain - your company's own context. That's what makes you hard, or easy, to catch. It's also the heart of an AI-native company.
We help you decide
At Elephant in the Room, we don't sell a religion - neither "all open" nor "all closed." The starting point is simple: not all your tasks have to run on the same model. We start from your real use cases and your company context. Where open-source models make sense - for cost, for control, or for data that must not leave your walls - we guide you through that transition: which models, local or cloud, for which tasks. And above all, we build the whole harness around it - router, policy layer, governance - so the thing actually holds up in production.
And that's how we run ourselves: our own company runs on a multi-model architecture, with no dependence on any single vendor. We don't sell a theory we don't practice.
The best model is the one you can swap tomorrow without breaking anything.
