Is it time to switch to open source models?

One of the most common questions I get right now is whether it is time to move to open source models. It is a fair question, because a year ago the honest answer was mostly no, and today it is genuinely "it depends."

Here is why it changed, and how I would think about the decision.

Open models are credible now

The gap has closed. When you look at recent releases, open models are roughly one step behind the frontier. If you are on the latest closed model, a good open model is often a version or so behind, which for a lot of work is more than enough.

The economics are the part that gets people's attention. Vercel, one of the big gateways for AI-powered code, is now running around 20% of its workloads on open source at about 3% of the cost. That is a real signal. Open source has moved from an experiment to something you can run in production.

Two mistakes to avoid

Once teams accept that, I see them fall into one of two traps.

The first is using a frontier model for every task. If you reach for the most capable model for simple, high-volume work, you are paying top rates for capability you do not need.

The second is the opposite. It is assuming that open or free always means cheapest. Once you account for hosting the model, running the infrastructure, and managing it over time, the cheapest sticker price is not always the cheapest total cost.

The way through both is to stop treating this as a single decision. It is a series of decisions, made task by task.

How I would actually make the switch

I would do it on a workload basis, not all at once. Pick a workload, test open models on your real inputs and outputs, and check them against the failure modes you already understand. Start where volume is high, where the work is easy to review, and where you can recover if something goes wrong. That is where the economics pay off first, and where a mistake is cheap to catch.

Then be honest about where to hold the line. Keep frontier models for the work that is genuinely complex, or where an error is expensive. In those cases you are paying for reliability, and it is worth it. The goal is to pay for that capability only where it converts into real value, rather than by default across everything.

The simplest version of the rule I use: open source wins when a task sits above the quality floor you need and below the cost ceiling you are willing to pay.

What this means if you sell to these teams

If you are in the technology channel, there is an opportunity here. The businesses you serve are going to need help making these calls, and the single most useful thing you can build for them is model portability. The best model for a job today may not be the best model in a few months. Something that needs a frontier model now could be an open source job by next quarter. If a customer's workflows are locked to one model, they cannot capture that. If they can switch models easily, they bank the savings over time.

The takeaway

Open source models are viable alternatives now, and for a lot of workloads they are the right choice on both quality and cost. But it is not automatic, and it is not all or nothing. Move workload by workload, keep frontier where the stakes are high, and build so you can change your mind later.

That is this week's episode of The AI Take. Sidney Minassian and I get into all of it, including the mistakes and the practical rule, in about eight minutes.

Watch the full episode on YouTube: https://www.youtube.com/watch?v=AXM87n3dpd0

If this was useful, subscribe to The AI Take newsletter for one practical AI lesson each week, and follow along on YouTube.

About 1KE:

1KE is the presales autopilot for technical sales teams. It turns discovery calls, emails, and technical documents into structured requirements and customer-ready solutions in minutes, so teams can respond while the deal is still warm.

Learn more → https://1ke.ai

Michael → linkedin.com/in/michaelchanter/ · https://x.com/1keceo

Sidney → linkedin.com/in/sidneyminassian/

Follow 1KE → linkedin.com/company/1ke-ai/