Despite this, businesses continue to deploy proprietary AI systems in production more frequently, noted Mozilla. According to the report, 73 per cent of large enterprises have adopted closed AI, compared with 57 per cent using open models.
If open AI has become nearly as capable as closed AI, why do organisations continue to favour proprietary systems? The answer lies less in the models themselves and more in everything required to deploy, manage and govern them.
Open-weight AI versus proprietary models
To understand Mozilla’s argument, it is important to distinguish between the two approaches.
Proprietary, or closed, AI refers to models whose weights, training processes and infrastructure remain under the developer’s control. Providers typically offer access through application programming interfaces (APIs) and subscriptions, with customers paying according to usage while the vendor manages infrastructure, updates and security.
With open-weight AI, developers release trained model weights, allowing organisations to download, modify and run the models on their own infrastructure. Companies such as DeepSeek, Alibaba, Mistral and Meta have popularised this approach.
The distinction has become increasingly important as enterprise AI adoption expands.
For organisations handling large AI workloads, running open-weight models internally can lower long-term costs, improve control over sensitive information and enable greater customisation. It also transfers responsibility for infrastructure, cybersecurity, software updates and compliance to the organisation.
Open-weight AI is closing the capability gap
According to the report, closed models led open-weight alternatives by 8.04 percentage points in January 2024. By August 2024, the gap had narrowed to 0.5 percentage points. By February 2025, the release of DeepSeek-R1 had brought it close to zero.
The gap has since widened slightly to 3.3 percentage points as reasoning-focused closed models pulled ahead again.
Mozilla found that open-weight models are at or near parity in coding, instruction-following and general knowledge. Closed models retain an advantage in reasoning, long-context retrieval and agentic tasks.
Costs have declined almost as quickly as capability has improved.
Mozilla found that running a model with performance comparable to GPT-4 cost about $20 per million tokens three years ago. Today, similar performance costs about $0.40 per million tokens, representing a 50-fold decline.
Tokens are the small units of text that AI models process.
The report attributes much of this price decline to open models. Meta’s Llama 3.1 and DeepSeek sharply reduced prices within a few months in 2024, forcing the wider market to respond.
Open models lead in tokens but not in requests
The cost decline is reflected in usage. Open-weight models accounted for nearly one-third of all tokens processed on OpenRouter by late 2025, and Mozilla’s data shows that the share continued to rise through June 2026.
By June, the five highest-volume models on the platform by token count were all open-weight: DeepSeek V4 Flash, Xiaomi’s MiMo-V2.5, Tencent’s Hy3 Preview, MiniMax M3 and a stealth model called Owl Alpha. Together, they processed more tokens than any closed model on the leaderboard, including Anthropic’s Claude family.
That token dominance, however, comes with an important qualification.
Open models are not handling more individual queries than closed alternatives. By request count, US-based closed-model providers continue to lead. Google accounts for 28.7 per cent of requests on OpenRouter, while Anthropic and OpenAI together account for about another 20 per cent.
Open models are disproportionately handling token-intensive workloads such as coding assistants and autonomous agents, where a single task may consume large volumes of tokens. Closed models continue to dominate everyday, request-led chatbot use.
Token volume and user popularity, therefore, measure different things.
Open AI is easier to adopt than deploy
Across Mozilla’s survey of more than 1,410 developers, 89 per cent of firms reported using at least one open-source AI component, while 79 per cent of developers said they used open models somewhere in their workflows.
Developer interest is not the main constraint. Production deployment is.
Only 51 per cent of organisations adopting open models successfully put them into production, compared with 63 per cent for closed models. The report attributes this gap to operational tooling and trust rather than model capability.
When organisations deploy AI with support from an established technology vendor, projects reach production 67 per cent of the time. When they attempt to build and deploy systems entirely in-house, the success rate falls to 33 per cent.
Mozilla also found that only 5 per cent of enterprise AI pilots, whether open or closed, currently deliver a measurable financial benefit.
Most organisations struggle to generate value from AI regardless of the type of model they use. Open models expose that challenge more clearly because they often lack a vendor providing infrastructure, support and implementation assistance.
Why company size helps closed AI more than open AI
Company size further illustrates the deployment gap. Among organisations that have begun using AI, the production success rate for closed models rises from 54 per cent at small companies, with two to 50 employees, to 73 per cent at enterprises with more than 1,000 employees.
For open models, company size has far less impact. Production success rises only from 53 per cent at small companies to 57 per cent at large enterprises. Mozilla concludes that deploying closed AI is often a problem that money can solve. Larger companies can purchase vendor support, managed infrastructure and implementation services.
Deploying open AI remains a problem that money alone cannot yet solve because the supporting ecosystem is less mature, regardless of the buyer’s budget.
Developers prefer open models on cost and privacy
Mozilla’s survey found that developers use open models across more use cases on average than closed systems: 5.1 use cases compared with 4.6. Among developers choosing open models, 30 per cent cited lower costs as the main reason, while 28 per cent cited privacy.
Between May and September 2025, open models handled about one-fifth of AI activity on the platform examined by Mozilla but generated only about 4 per cent of revenue.
For broadly comparable output quality, closed models charged nearly six times more per request.
Mozilla estimates that businesses could collectively save $24.8 billion annually by shifting more suitable workloads to lower-cost open models.
Why deployment tools matter as much as AI models
Mozilla also identifies a layer above the model itself as a reason open AI struggles in production: the harness.
A harness is the surrounding software that allows a raw AI model to take actions, retain context and operate safely within a business rather than merely answer questions. Independent harnesses once outperformed AI laboratories’ own tools by a wide margin. In one case, an externally developed harness beat a laboratory’s in-house tool by almost 22 percentage points on a widely used coding benchmark.
Within two months, however, the laboratory improved its harness to work specifically with its latest model. The in-house tool then overtook the best external alternative.
Mozilla’s conclusion is that organisations controlling both the model and its surrounding toolkit gain an advantage.
The report also identifies an unresolved challenge across the wider AI industry: there is no agreed standard governing when an AI agent should be permitted to take a serious, difficult-to-reverse action, such as transferring money or changing an official record.
Until that issue is addressed, it may remain a greater barrier to enterprise trust than benchmark performance.
Open-weight AI is becoming a geopolitical contest
Mozilla argues that the rise of open-weight AI is no longer merely a technology trend. It has become a matter of national strategy.
Chinese open models from DeepSeek, Xiaomi, Tencent and MiniMax occupy four of the five leading positions by token volume on OpenRouter.
The report links this rise to deliberate industrial policy. Releasing capable open weights allows Chinese companies to build global developer dependence while shifting computing costs to users’ hardware, partly offsetting the effects of semiconductor export controls.
Mozilla also cites an example of why organisations using closed frontier models are becoming concerned about optionality.
In June 2026, Anthropic’s Claude Fable 5 and Mythos 5 models became unavailable for about 20 days after a US export-control order restricted access.
Organisations relying on those endpoints had no control over the decision or recourse while access was suspended. The episode demonstrated that the risks associated with dependence on a closed provider are no longer theoretical.
Europe, meanwhile, is treating open weights as an issue of digital sovereignty rather than merely cost.
European Commission President Ursula von der Leyen has said Europe “cannot afford to depend on others for the technologies that keep our hospitals running, our energy grids stable and our services secure”.
India’s open AI strategy focuses on infrastructure
India’s approach differs from those of China and Europe. Mozilla says India’s strategy is less focused on building a single frontier model and more on creating the infrastructure needed to participate at scale.
The report cites 38,231 graphics processing units (GPUs) empanelled under the IndiaAI Mission at subsidised rates of about Rs 65 an hour, roughly 40 per cent below market rates.
The programme is backed by an outlay of Rs 10,372 crore and includes plans for 600 data laboratories.
India added more than five million GitHub developers in 2025, taking its total developer base to 21.9 million. According to the report, India has the world’s fastest-growing developer population and accounts for 13.6 per cent of DeepSeek’s monthly active users, second only to China.
Enterprise AI contest shifts beyond model capability
Mozilla concludes that the model-capability race is approaching parity. The more important contest is shifting to deployment tools, portable memory, permission standards and control over the infrastructure enterprises trust.
Whether open-weight AI can close this gap as effectively as it narrowed the capability gap will depend less on the next benchmark score and more on whether its supporting ecosystem matures.
Support, governance, security and accountability will have to catch up with what open models can already do.
