Open-weight AI models now handle 63 % of token traffic on OpenRouter, up from under 5 % just two years ago. The surge follows a 50-fold drop in per-token cost and a narrowing performance gap that is forcing developers to rethink their dependence on closed-source APIs.

The numbers behind the shift

Mozilla’s traffic analysis shows open-weight models have become the dominant source of compute on the OpenRouter platform. Token traffic counts every piece of text processed by a model, making it a reliable proxy for real-world usage. Jumping from sub-5 % to 63 % in such a short span signals a structural change rather than a fleeting trend.

Why the economics flipped

In 2023, running a high-performance model cost roughly $20 for every million tokens processed. Today that price is $0.40—a 50-times reduction achieved over the past 36 months.

Performance gap shrinks

Open models used to lag behind proprietary counterparts by about 8 % on standard benchmarks. Recent measurements put that gap at 3.3 %, a reduction achieved in just a few months through better data curation and architectural tweaks. The gap is now small enough that many developers see little downside to swapping a closed model for an open one, especially when cost is decisive.

Developers are moving en masse

Mozilla data shows 79 % of developers adding AI features now start with an open-weight model. The choice isn’t just about price; open models give teams full visibility into weights, licensing, and deployment options, which translates into faster iteration cycles.

Sovereignty adds urgency

Forty-seven countries have recently restricted processing critical workloads with foreign AI services. Open models can be hosted on domestic infrastructure, letting those nations retain data control and comply with emerging regulations. Policy pressure is accelerating adoption in finance, healthcare, and public administration.

The real moat is the “harness”

Mozilla’s analysis also highlights that surrounding tooling—prompt-engineering frameworks, safety filters, orchestration pipelines—often yields larger performance gains than swapping to a marginally better model.

Risks of staying locked into closed APIs

  • Vendor risk – Policy changes, pricing revisions, or government orders can instantly cut off access, leaving services offline until a new contract is negotiated.

A pragmatic stack for today’s developers

  • Default to open-weight models for the bulk of production workloads.
  • Invest in the harness: build evaluation suites, develop reusable prompt libraries, and integrate safety layers.
  • Design for portability: abstract model calls behind an interface that can route to multiple providers without code changes.
  • Reserve closed models for the roughly 10 % of tasks that demand cutting-edge reasoning or proprietary data that open models cannot yet match.

What to watch next

Open-weight models now dominate the AI token market because they are dramatically cheaper and almost as capable as closed alternatives. Teams that build a strong harness and keep their stacks portable will reap cost savings while insulating themselves from vendor and policy shocks. Closed models survive as specialty tools, but the bulk of AI work is already moving into the open.