Open-Source AI Is 6x Cheaper at Near-Parity, If You Can Pay the Production Tax

The thing the internet still argues about as an underdog cause has, by volume, already won. Open-weight models, the ones whose weights you can download and run yourself, now carry the majority of production tokens on OpenRouter, the marketplace that routes API calls across many model providers.

That is the finding at the center of Mozilla's State of Open Source AI Report, published in July 2026. The team debate on your timeline is about a race that, for a large share of real traffic, is already over.

So the interesting question is not which side to root for. It is what to do about it if you are shipping something. The honest answer is not "switch everything to open." It is a decision procedure, one workload at a time, built to survive a gap that will not sit still.

Is open-source AI actually cheaper than closed models?

Yes, at the API level, and by a wide margin. As of Mozilla's July 2026 report, closed models cost roughly six times more per call than open models at comparable capability, and the measured capability gap has fallen to 3.3%. Two years earlier it was 8%. The premium for closed is now mostly a premium for convenience.

The broader price story is the same shape. Inference, which just means running a model to get an answer back, has fallen about 50x in 36 months, from around $20 to $0.40 per million tokens as of the same report.

Mozilla puts the estimated unrealized savings from the open-versus-closed price asymmetry at roughly $24.8 billion a year. That last figure is an estimate built on pricing assumptions, not a receipt, so hold it loosely.

The direction is not in doubt.

This is the same instability we wrote about in why token prices keep getting repriced: today's price reflects today's financing and today's leader, not a settled cost of the work.

So the naive read is "move everything to open and pocket the six times." That is where most hot takes stop. It is also wrong, because the token price is not the cost.

What does it actually cost to run an open-source LLM in production?

More than the sticker price implies. In the same Mozilla report, only 51% of open-model teams reach production, versus 63% for closed. A third of open-model projects stall before they ship, and the barriers are named and sized rather than hand-waved:

• Infrastructure cost (27%)
• Security and compliance (26%)
• Maintenance (24%)
• Deployment complexity (23%)

Read that list again, because it is the whole argument. Every one of those line items is work a closed API was quietly absorbing for you.

When you rent a model by the token, the vendor eats the servers, the patching, the uptime, the compliance surface. When you run the weights yourself, you inherit all of it.

This is the production readiness gap we keep coming back to. The demo is not the deadline, and getting to production is its own discipline.

Open is cheaper per token and more expensive per unshipped project.

That reframe is the entire trade. The six-times saving is real, but it is conditional. You capture it only if you can pay the production tax, and the tax is exactly the operational envelope a closed vendor hides in the price. The teams that win treat the model as the cheap part and the envelope around it as the real investment.

When should you use an open-source model instead of a closed one?

When the task is well-scoped, high-volume, and cost-sensitive, and when you can carry the operational load, open models are now the rational default. When the job needs the hardest reasoning or long-context retrieval, or when a 3.3% capability difference decides the outcome, closed still earns its premium. The mature answer is almost never all-or-nothing.

The decision is per-workload, not per-company. A support-ticket classifier running millions of times a day and a frontier research agent doing multi-step reasoning are not the same bet. Nothing says they belong on the same model.

The move in practice: send each workload to the cheapest model that clears its quality bar, measure that bar with tests you own, and revisit as prices shift. That is unit-economics thinking applied to model choice, the same lens as metered versus included pricing. Not a platform decision you defend forever.

The team-sport framing hides all of this. It wants you to pick open or closed the way you pick a jersey. There is no jersey that is cheapest-at-quality for every workload, and the moment you commit to one for all of them, you start overpaying on the jobs that needed the other.

How do you avoid getting locked into one AI model?

Design for portability from the start. The point is not to predict the winner. It is to make the winner not matter very much, so that switching models is a config change instead of a rewrite. Three habits do most of the work:

• Build against a compatible API surface, so swapping models is a configuration change.
• Keep a tested fallback model wired in, not a diagram of one.
• Never hard-code one vendor's quirks into a product-critical path.

This matters now specifically because the gap is not stable. The capability lead changes hands in weeks, not years.

Open-weight models like GLM 5.2 have landed frontier-adjacent coding at roughly $4.40 per million tokens, a fraction of frontier pricing, reachable through compatible endpoints at near-zero switching cost. Others, like Inkling and Kimi K3, are days apart on the calendar. Whatever is cheapest-at-quality for a given job today is probably not cheapest-at-quality for it next month.

That mobility is the real reason portability beats loyalty. It is the same lesson as the harness lock-in trap: the thing that traps you is rarely the model, it is everything you welded to one model.

The anti-pattern is picking a side and marrying it. All-in on one closed vendor is the lock-in trap. All-in on self-hosted open is paying the full production tax on workloads that never needed it. Both convert a routing decision into a structural bet, and structural bets lose to a gap that keeps moving.

Portability is not free either. It costs you an abstraction layer, a test harness per model, and some feature loss when you refuse to depend on any one vendor's best trick. That price is worth paying in exactly one condition: when the gap is small and mobile. Which is to say, now.

When one model is decisively ahead for your workload, pay for it and move on. When the field is 3.3% apart and the lead trades hands every few weeks, stay liquid.

The war is the wrong frame

There is a version of this moment that reads as a victory lap for open source, and a version that reads as a warning that open stalls in production. Both are half the report. The value is in holding both numbers at once: cheaper per token, more expensive per shipped project, and a lead small enough to cross on purpose.

The open-versus-closed war is the wrong frame for anyone actually building. The numbers say the gap is small, cheap to cross, and constantly moving. The builders who come out ahead this year are not the ones who bet the right side. They are the ones who built so they never had to bet at all.