Paying for a frontier AI model now buys you about four months of advantage

Key points

• As of September 15, 2025, Mozilla's State of Open Source AI report found the performance gap between top US frontier models and the best open-weight models (free AI models whose underlying code anyone can download and run) has closed to just 4.4 months.
• Moonshot AI's Kimi K3, an open-weight model, scores only three points below Anthropic's Fable 5 on a standard AI performance index, at 30 percent of the cost.
• Mozilla's CTO says the decision to pay for a closed, proprietary model should depend on the specific task, not on the organisation as a whole.
• Most organisations, Mozilla says, should treat open models as the default for the bulk of their work.

For years, if you wanted the best AI, you paid for the best AI. The gap between costly closed models from big US labs and free, downloadable alternatives was wide enough to justify the bill. That gap has nearly vanished.

Mozilla's latest State of Open Source AI report, published on 15 September and first reported by Ars Technica, puts the performance difference at just 4.4 months. In practical terms: the best free model available today performs roughly as well as the best paid model did about four months ago.

What does that actually mean for your organisation?

For most routine work, open-weight models are now good enough, and the cost difference is hard to ignore. Moonshot AI's Kimi K3, an open-weight model anyone can download, scores within three points of Anthropic's Fable 5 on the Artificial Analysis Intelligence Index, a standard benchmark measuring overall AI capability. Kimi K3 costs about 30 percent of what the Anthropic model charges per use.

Mozilla's chief technology officer, Raffi Krikorian, was direct about the conclusion. "Closed earns its premium in a few places: expert professional work and long context," he told Ars Technica by email. "We see the decision to pay for closed as workload-specific rather than organization-specific."

A legal team drafting complex contracts or a researcher sifting through thousands of documents may still get genuine value from premium models. A shop owner writing product descriptions or a teacher building a quiz probably does not.

The Mozilla report recommends that most organisations treat open models as the default, then reach for a paid frontier model only when a specific task genuinely demands it. We've tracked the accelerating pace of new model releases since early September, and the competitive pressure on closed-model pricing is only building.

What should ordinary users watch for?

This shift matters beyond IT departments. As companies quietly swap expensive models for cheaper alternatives to cut costs, the AI tool you used last month may be running on different software today, with no announcement. Performance on your specific use case could improve or slip.

Ask your vendor which model sits behind the product you rely on. If the answer is vague, that's a problem. The benchmark scores Mozilla cites measure average capability; they don't guarantee the model handles your particular task well.

My read: the closed-versus-open debate is settling into something more useful than a hunt for a single winner. It's a question of matching the right tool to the right job, and organisations that figure that out early will spend less and get comparable results for most of what they do.

Common questions

Are open-weight models safe to use for sensitive business data?

Can a small business actually run one of these open models?

Running a large open-weight model yourself requires specialist hardware and technical staff. Most small businesses are better served by a cloud provider that hosts open models on their behalf, which keeps costs low without needing an in-house AI team.