Every team building with AI eventually faces the same fork in the road: use a closed model from a provider like OpenAI, Anthropic, or Google through an API, or run an open-weight model such as Llama or a strong open alternative on your own infrastructure. In 2026, the honest answer is that this is no longer a question about which is smarter. It is a question about fit.

The capability gap has narrowed

The most important change is that open models caught up on ordinary work. By 2026, open-weight models reach roughly 85–90% of closed-model performance on most standard enterprise tasks, and the gap on general knowledge benchmarks has effectively closed.

Where closed models still lead is the hard edge: complex multi-step reasoning, novel coding problems, and nuanced research synthesis. If your product depends on the most difficult 10% of tasks, that edge is real and worth paying for. If it does not, you may be paying for capability you never use.

The cost difference is large

This is where the decision usually gets settled. Industry comparisons in 2026 put closed-model APIs at roughly $10 per million input tokens, while comparable open models run at around 60 cents per million when self-hosted — with open models broadly closing 70–90% of the capability gap at five to ten times lower per-token cost.

Two cautions. First, prices move constantly, so treat these as indicative rather than fixed. Second, self-hosting is not free: you are trading a per-token bill for engineering time, GPU capacity, and operational responsibility. Below a certain volume, the API is genuinely cheaper once staff costs are counted.

Closed models: the practical case

Choose closed when:

  • You need the strongest available reasoning for hard problems.
  • You want to ship quickly without managing infrastructure.
  • Your volume is low or unpredictable, making pay-as-you-go efficient.
  • You value automatic access to each new model generation.

The trade-offs: higher per-token cost at scale, dependence on a vendor's pricing and roadmap, and sending data to a third party.

Open models: the practical case

Choose open when:

  • You run high volume where per-token savings compound significantly.
  • Your data cannot leave your environment — healthcare, defence, regulated finance.
  • You need deep fine-tuning on domain-specific data.
  • Vendor lock-in is a genuine strategic risk for your business.
  • You need a model version that stays stable and will not change beneath you.

The trade-offs: you own the operational burden — serving, scaling, monitoring, security patching — and you carry responsibility for safety behaviour that a provider would otherwise handle.

What most teams actually do

The framing of "open versus closed" is increasingly outdated in practice. Most serious production deployments in 2026 use both: closed frontier models for general-purpose and hard reasoning tasks, and fine-tuned open models for high-volume, cost-sensitive, or domain-specific workflows.

That hybrid pattern is exactly the routing logic we described in the 2026 AI price war — send each task to the cheapest model that clears your quality bar, and escalate only when the harder model changes the outcome. The same discipline applies inside AI agents, where long chains of steps multiply every unnecessary cost.

How to decide, in order

  • Start with constraints. If data cannot leave your building, the decision is already made.
  • Estimate real volume. Low or spiky volume favours APIs; steady high volume favours self-hosting.
  • Test on your own tasks. Public benchmarks are a poor proxy for your specific workload.
  • Count total cost. Include engineering time, not just the per-token price.
  • Keep the door open. Design your code so swapping models later is not a rewrite.

Key takeaways

  • Open models now reach roughly 85–90% of closed-model performance on most enterprise tasks.
  • Closed models retain an edge on complex reasoning, novel coding, and nuanced synthesis.
  • Open models can cost five to ten times less per token, but self-hosting adds engineering and operational costs.
  • Regulated, air-gapped, or heavily fine-tuned use cases favour open weights; speed and peak capability favour closed.
  • Most production teams use both, routing each task to the cheapest capable option.

The bottom line

Stop asking which type of model is better and start asking which is better for a specific job. In 2026 the winning strategy is not loyalty to open or closed — it is building systems flexible enough to use whichever one fits each task, and to change your mind as prices and capabilities shift.