IA · 15 September 2026 · 5 min read
Frontier Model Lead Narrows to Four Months: Mozilla's Report and Salesforce's Open-Weight Shift
In brief: According to Mozilla's latest report, the performance gap between Silicon Valley's proprietary frontier models and leading open-weight alternatives has closed to just 4.4 months, despite closed models costing up to five times more. Concurrently, Salesforce unveiled Koa at Dreamforce—a reasoning model built on Nvidia's open-weight Nemotron for Agentforce—highlighting a decisive structural pivot across enterprise IT toward open and sovereign architectures.
by Team Mocchi's
For months, the prevailing narrative from frontier AI labs has maintained that only multi-billion-dollar proprietary infrastructure could deliver the reasoning capabilities required for complex enterprise workflows. However, data from Mozilla's State of Open Source AI report, previewed by Ars Technica, reveals a starkly different reality: the performance head start held by closed Silicon Valley models over top open-weight counterparts has fallen to just 4.4 months.
The cost disparity, on the other hand, remains significant. Paying for access to closed frontier systems currently costs roughly five times as much as deploying open-weight solutions. Leading open models, such as Moonshot AI's Kimi K3, now trail top proprietary benchmarks like Anthropic's Fable 5 by only three points on the Artificial Analysis Intelligence Index, while operating at merely 30 percent of the cost.
A price premium increasingly hard to justify
The narrowing delta is calculated using evaluations from the METR research group, which tracks model time horizons: the length of complex human tasks an AI agent can execute autonomously with a reliable 50 percent success rate. While closed frontier systems can still handle tasks approximately 1.7 times longer than the top open weights, the cadence at which open alternatives close that gap has accelerated steadily throughout 2026.
As Mozilla Chief Technology Officer Raffi Krikorian pointed out, paying a premium for closed frontier APIs is becoming a task-specific rather than an organization-wide decision. Enterprise workloads that genuinely require closed frontier systems are narrowing down to deep specialized reasoning, extreme context windows, and high-density retrieval. For routine workflows, organizations like DoorDash have already migrated daily volume to open weights, reserving costly frontier calls strictly for edge cases.
The enterprise pivot: Salesforce and Nvidia introduce Koa
Evidence of enterprise buyers stepping back from frontier API lock-in was underscored this week at Dreamforce. As reported by TechCrunch, Salesforce launched Koa, its first reasoning model tailored for sales, marketing, and customer service operations within its Agentforce suite.
Rather than turning to closed providers, Salesforce partnered with Nvidia to post-train Koa on top of the open-weight Nemotron foundation. By doing so, Salesforce addresses key corporate requirements: eliminating data leakage into third-party labs, adhering to strict corporate governance, and optimizing token efficiency by aligning the reasoning engine strictly with enterprise workflows rather than generalized benchmarks.
While closed frontier providers still offer turn-key compliance packaging and direct accountability—which matters to enterprises without dedicated in-house machine learning engineering teams—the widespread adoption of intelligent routing layers makes relying solely on one proprietary provider obsolete.
Mocchi's take
For European businesses and software engineering teams, Mozilla's findings and Salesforce's Koa represent an essential validation of architectural independence. Funneling enterprise workloads blindly into costly proprietary foreign APIs creates unsustainable operating expenses and unnecessary data governance exposure under the EU AI Act. The strategic imperative for companies today is building hybrid orchestration pipelines: hosting domain-tuned open-weight models within private or sovereign infrastructure for core tasks, while reserving commercial frontier APIs strictly for high-entropy exceptions.