Paying for Closed AI Now Means Buying a 4-Month Lead at 5x the Cost, Mozilla Study Finds

In many development environments, model selection is no longer a company-wide either/or decision about whether to sign a blanket contract for the top-performing service. As Mozilla CTO Raffi Krikorian points out, the decision to pay a roughly 5x cost premium for a commercial closed model has entered a phase where it should be made case-by-case based on the nature of individual workloads, rather than as a blanket organizational policy.

For the vast majority of everyday development and routine tasks, open-weight models already offer sufficient accuracy along with overwhelming price competitiveness. Meanwhile, the narrow band of work where expensive proprietary models still deliver value commensurate with their cost is increasingly limited to tasks requiring deep specialized expertise, dense information retrieval, or extremely long context windows.

{
  "type": "bar",
  "title": "Cost-per-Task Comparison (Terminal-Bench 2.1 Evaluation)",
  "unit": "USD/test",
  "categories": ["GLM 5.2 (Open)", "Claude Opus 4.8 (Closed)"],
  "series": [
    { "name": "Cost per run under neutral harness", "values": [0.89, 2.82] }
  ],
  "caption": "Measured completion cost on the Terminal-Bench 2.1 benchmark using Vals AI's neutral harness. Under identical harness conditions, the open model recorded a cost less than a third that of the closed model.",
  "source": "Vals AI / Mozilla State of Open Source AI v1.1"
}

Behind this cost disparity lies the fact that open-weight models have rapidly closed in on the top-tier commercial offerings. A generational gap once thought to span six months to nearly a year has now compressed to little more than a quarter. Why do companies keep paying for expensive API usage, and for which kinds of work can they immediately switch to cheaper open models? Drawing on published measurement data and benchmarks, this article examines the practical boundary lines for model selection.

The 4.4-Month Gap and the 8-to-12-Hour Line That Divides Model Choice

The "4.4-month" lead time calculated by Mozilla is derived from its own regression analysis based on raw data collected by METR (Model Evaluation and Threat Research), a nonprofit AI evaluation organization. An independent estimate using the Epoch Capabilities Index (ECI), which tracks AI progress via a composite score, also puts the gap between closed and open models at roughly 4 months (an average 8-point ECI difference, with a 90% confidence interval of 7 to 11 points)—a conclusion that broadly aligns across different calculation methods.

Based on this objective data, Mozilla's report explicitly recommends that, for most organizations, open models should be the default choice for the large majority of everyday tasks. However, this does not mean the gap between the two has disappeared entirely for every task. According to Krikorian's explanation to Ars Technica, closed models can still justify their premium pricing in areas requiring highly specialized professional-level judgment, dense information extraction, and the ability to reliably sustain extremely long context.

Mozilla itself attaches several caveats to this claimed advantage. Open model developers tend to over-optimize for public benchmark scores, meaning the actual gap in real-world overall capability may not have narrowed as much as the numbers suggest. In addition, unreleased models that major AI labs are developing internally but have not yet made public are not included in the comparison baseline. As a result, the real distance to the actual frontier may be somewhat wider than the statistical 4.4 months suggests.

As this comparison table shows, the gap in scores on major benchmarks has already narrowed to a minimal degree. Yet the reason commercial models are still not abandoned in practice lies not in raw capability differences, but in the limits of task duration and reliability discussed below.

What Task-Duration Math Reveals: The 1.7x Task Length Still Monopolized by Commercial Models

METR pioneered an approach to measuring capability gaps not through abstract test scores, but through the physical time required to complete real work. METR defines a concept called the "50% time horizon." This refers to the length of task—measured in the time a human expert would need to complete it—that an AI model can successfully finish end-to-end with a 50% probability. This metric is calculated from data drawn from practical benchmark suites such as RE-Bench and HCAST, which cover complex software engineering challenges, cybersecurity, and scientific reasoning. METR itself notes, however, that its estimates of human task time may include some overestimation.

In research published by METR in March 2025, the length of tasks AI could autonomously solve was found to double roughly every seven months, with signs that this pace had accelerated further since 2024. When Mozilla applied this methodology to the latest data, it found that closed models are doubling their task length every 5.5 months, while open models are doing so at an even steeper pace of 3.9 months.

{
  "type": "bar",
  "title": "Comparison of Task Duration Sustained at 50% Success Rate",
  "unit": "hours",
  "categories": ["Open Frontier", "Closed Frontier"],
  "series": [
    { "name": "Current models (measured)", "values": [7, 12] },
    { "name": "Projected in 4 months (Mozilla estimate model)", "values": [12, 20] }
  ],
  "caption": "Comparison of sustained task duration based on METR's 50% time horizon methodology. Closed models currently handle tasks about 1.7x longer than open models, but in four months open models are expected to reach today's closed-model level.",
  "source": "Mozilla State of Open Source AI v1.1 / METR raw data"
}

This difference in doubling cycles defines the concrete work limits of each camp at present. According to the practical model Krikorian presents, the top closed model can handle tasks about 1.7 times longer (1.74x, per Mozilla's regression line fitting) than the longest tasks the top open model can reliably complete. To put it concretely: when the open frontier can handle work equivalent to 7 hours of a human expert's time, the closed frontier can handle 12 hours.

After four months pass, open models will be able to handle the 12-hour tasks that once marked the closed-model ceiling, while closed models will have advanced into the 20-hour range. From this, the practical guidelines for dividing work—within this comparison and the 50% time horizon framework—can be organized into three categories:

1. Tasks taking a human expert under 8 hours: Both open and closed models can reliably complete these, and given cost considerations, they should be handed to open models.
2. Tasks taking a human expert 8 to 12 hours: This is currently the domain where top-tier closed models can typically complete the work with over 50% reliability, while open models often still cannot.
3. Complex tasks taking a human expert over 12 hours: With current AI models, it is generally difficult for any model to reliably complete these with over 50% probability.

In other words, the justification for companies relying on expensive commercial closed models is usually concentrated in this narrow "8-to-12-hour" band of work. For anything shorter, open models can keep pace well enough; for anything longer, no model—regardless of type—can currently be expected to complete the work reliably.

However, a major caveat looms over this comparison of task length. Many AI companies offering frontier models do more than simply rent out API access to a base model. They design specialized software execution environments—known as "harnesses"—tailored to their own models, enabling the model to call various tools and memory and iterate autonomously through trial and error. Commercial models running on vendor-designed, purpose-built harnesses perform far better than when run on generic third-party frameworks. It is not easy to disentangle whether a given comparison reflects the intelligence of the model itself or the sophistication of the surrounding support software.

What a Neutral Environment Reveals About Cost Efficiency: A 5x Gap in Monthly Spend

Efforts are also underway to strip away the performance boost from vendor-built harnesses and measure model capability on a level playing field. Benchmark firm Vals AI ran "Terminal-Bench 2.1," connecting models from various companies to its own neutral, single harness environment to evaluate their ability to autonomously execute coding and systems-operations tasks.

The results shocked the developer community. "GLM 5.2," an open-weight model released by China's Zhipu AI (branded Z.ai), achieved a 67.79% success rate under its maximum reasoning-resource configuration. That is a mere 0.75-point gap from the 68.54% recorded by Anthropic's top commercial model, "Claude Opus 4.7"—effectively putting it on par in capability.

Even more striking was the difference in execution cost. When measuring the average cost of completing a single test, Claude Opus 4.8 required $2.82, while GLM 5.2 completed the same task for just $0.89—a gap of about 3.17x per completed task. It should be noted that Mozilla's report and Ars Technica's coverage also reference an overall estimate that closed models are "roughly 5x more expensive per task" (or open models roughly 5x cheaper), though the specific absolute dollar figures underlying this approximately 5x ratio are not disclosed in the published materials.

This picture of price disruption is corroborated by other independent benchmarks. On the composite index maintained by Artificial Analysis, which conducts comprehensive comparisons of AI model performance (the Intelligence Index v4.1.1, dated September 1, 2026), China's Moonshot AI released an open model, "Kimi K3," that scored 60 points. That places it just 1 to 3 points behind the top-tier group scoring around 61, which includes Anthropic's next-generation closed model "Fable 5." Yet Kimi K3's per-input-token usage price is set at roughly 30% of Fable 5's.

Handling these figures also requires caution. In Artificial Analysis's evaluation, commercial frontier models such as Fable 5 and Sol are tested as "complete operational systems" incorporating guardrails, automatic fallback mechanisms, and safety filters. Open-weight models, by contrast, are often measured as "bare models" without such added features. This is a comparison of pure output performance that disregards system-level robustness or the presence of compliance mechanisms, and that context deserves consideration.

Also, on the Epoch Capabilities Index maintained by AI research organization Epoch (data as of September 1, 2026), Kimi K3 recorded 157 points, closing to within 5 points of Claude Fable 5 and Opus 5's 162. Epoch's analysis concludes that the average 8-point gap observed between the top commercial camp and the open camp since January 2026 translates to roughly a 4-month development lag.

Explosive Traffic and 4% of Revenue: The Economic Distortion Facing Open Models

This rapid catch-up in performance and price disruption is significantly transforming actual developer usage patterns. Usage statistics from OpenRouter—a router service that unifies API access across a wide variety of models—show that as of August 2026, open-weight models accounted for 8 of the top 10 models by token consumption. This reveals how, for repetitive tasks such as everyday code generation, document summarization, and data structuring, engineers are unhesitatingly choosing cheaper open models.

A hybrid strategy adopted by U.S. food delivery giant DoorDash exemplifies this trend. The company uses Moonshot AI's Kimi for internal routine work and high-volume customer support processing to minimize infrastructure costs, while reserving Anthropic's pricier Fable models exclusively for high-difficulty analytical tasks and incident response that would require extended deliberation even from human experts.

Yet despite the explosive growth in usage volume, a serious distortion remains in the revenue structure of the model-provision business. A working paper by Harvard Business School's Frank Nagle and Daniel Yue, written for the Linux Foundation and titled "The Latent Role of Open Models in the AI Economy" (published November 2025), reports a starkly imbalanced market reality.

Despite commercial closed models charging on average roughly 6 times the usage price of open models, they nonetheless dominate about 80% of overall market usage share and, remarkably, 96% of total market revenue. Open models' share of revenue, by contrast, stood at just 4%.

However, the Nagle and Yue data analyzes OpenRouter transaction records from May through September 2025, covering less than about 1% of total global LLM API spending. As Krikorian himself points out, given the literally explosive growth in open model usage and performance over the past year, the current revenue share has likely shifted somewhat toward the open side.

In the same paper, Nagle and Yue estimate that if companies stopped uniformly relying on commercial models and instead properly reallocated their workloads to open models, they could achieve annual cost savings of $20 billion to $48 billion. According to their baseline estimate, extrapolated from Menlo Ventures' 2025 market-size projections, an estimated $24.8 billion (about ¥3.7 trillion) in economic value per year remains unrealized simply because the shift to open models has not progressed further.

The reason companies, despite recognizing this enormous cost-saving opportunity, still pour the majority of their spending into commercial closed models is straightforward. Commercial models come bundled from the outset with the ease of a single-contract deployment, rigorous enterprise-grade compliance packages, service-level agreement (SLA) guarantees, and vendor accountability. Most ordinary companies critically lack the engineering staff needed to securely and reliably host and maintain massive open-weight models on their own servers or private cloud.

Also not to be overlooked is the fact that many of the Chinese-made models currently labeled "open" are not, strictly speaking, open-source software. While the weight parameters themselves can be freely downloaded and used by anyone, the specific training datasets showing what the model actually learned, the details of preprocessing pipelines, and the source code used for training remain hidden inside a black box.

Concentration in Chinese Capital and an "Android Replay": The Decision-Making Discipline Companies Need

The fact that the vast majority of open-weight models currently benefiting developers worldwide originate from Chinese labs raises new geopolitical and structural questions. Krikorian draws a parallel to the strategy once taken by giant U.S. tech companies in the mobile market: giving away the Android operating system for free worldwide while seizing control of the entire surrounding ecosystem of APIs, app stores, and cloud services. He sees an eerily similar structure being replicated at the foundational layer of generative AI.

What concerns Krikorian is not so much that model developers are based in China per se, but the excessive concentration of power within the technology ecosystem overall. The current situation has fallen into a bipolar trap: "the world's best closed models are concentrated among a handful of U.S. companies, and the world's best open models are concentrated among a handful of Chinese companies." To avoid a scenario in which a single country ends up dictating the default specifications for AI worldwide, public institutions and research organizations in the U.S. and Europe must genuinely commit to building an open competitive framework.

The fact that the diverse ecosystem surrounding open-weight AI is, in practice, substantially underpinned by Chinese capital reveals a precarious equilibrium in which diversity and regional concentration are occurring simultaneously. As history shows with Linux, which has become indispensable social infrastructure today, maintaining a shared commodity layer requires a coalition of stakeholders independent of any single for-profit company's interests.

What Mozilla advocates is the formation of an "alternative coalition" driven not by market logic but by public mission. Specifically, it envisions a framework built on four pillars:

1. Contributing public compute resources: Investing national-scale supercomputing resources to nurture fully open reference models. The "Apertus" project led by Switzerland's national supercomputing center serves as a leading example.
2. Leadership by neutral nonprofit foundations: Just as neutral parties protected the open protocols of the Web in the early days of the internet, neutral bodies should establish and maintain standards for model specifications, execution harnesses, and autonomous agent communication.
3. Participation by operating companies that benefit from commoditization: Ordinary companies that can dramatically cut infrastructure costs through the spread of cheap open models should directly participate in funding and developing such a consortium.
4. Philanthropic support for audit infrastructure: Philanthropic funds should focus grants on neutral third-party model evaluation functions and safety/bias audit infrastructure—areas for-profit companies tend to avoid investing in.

At the same time, Krikorian warns that unless the industry moves beyond the current practice of merely distributing model weights and shifts toward a genuine open-source approach that discloses training data and validation procedures in full, companies cannot entrust AI with decisions core to their business. Systems whose training data biases cannot be audited will fundamentally fail to earn trust in fields such as finance, healthcare, and critical infrastructure.

The "pay rule" for model selection that Mozilla's report ultimately offers to practitioners is highly pragmatic. Paying a roughly 5x premium per task for a closed model is justified only when the work falls into the difficult 8-to-12-hour band for a human expert, and when there is a pressing deadline that cannot wait for the "four months it will take the open camp to catch up."

There is no longer any necessity to assign expensive proprietary models to routine internal tasks that will keep running unchanged into the next quarter. By the time that quarter has passed, open models will be able to replicate at a far lower cost the capabilities currently monopolized by top-tier models. Calmly assessing the range of task durations one's own workloads demand, and using the four-month time gap as a yardstick for cost-effectiveness, will be what determines the success of AI infrastructure investment going forward.