One of the most consequential divides in the AI industry is not between any two specific companies, but between two fundamentally different approaches to releasing AI models: keeping them proprietary and locked behind a paid API, or releasing their weights openly for anyone to download, modify, and run. This split shapes everything from pricing and data privacy to how quickly the overall field advances, and the question of which approach is actually “winning” depends heavily on how you define victory.
Defining the Two Camps
Proprietary models, like OpenAI’s GPT-4 and GPT-4o, Anthropic’s Claude family, and Google’s Gemini, are only accessible through the company’s own hosted API or chat interface. Users and developers send requests to the company’s servers and receive responses back, but they never gain access to the underlying model weights, the vast set of numerical parameters that actually encode what the model has learned. Open-weight models, including Meta’s Llama family, Mistral’s open releases, and China’s DeepSeek models, make those weights publicly downloadable, usually under a license permitting commercial use, allowing anyone with sufficient hardware to run the model entirely on their own infrastructure without depending on the original company’s servers at all. It’s worth noting that “open-weight” and “open-source” are not quite the same thing, since most of these releases do not include the full training data or code used to create the model, only the final trained weights themselves.
Where Proprietary Models Still Lead
On raw capability at the very top end, proprietary models have generally maintained a lead, at least for brief periods following each major release. OpenAI’s most capable models, Anthropic’s top-tier Claude models, and Google’s most advanced Gemini variants have typically topped independent benchmark leaderboards and human preference rankings like Chatbot Arena when first released, reflecting the enormous computing resources and specialized research talent that the best-funded closed labs can direct toward a single flagship model. These companies also generally offer more polished consumer products, more extensive safety testing and red-teaming before release, and dedicated customer support for enterprise deployments, advantages that come more easily to well-funded companies focused on commercial products rather than research releases intended primarily for the developer community.
Where Open-Weight Models Have Closed the Gap
Despite this historical lead, the gap between the best open-weight models and the best proprietary models has narrowed dramatically faster than most industry observers expected just a few years ago. Meta’s Llama 3 and 3.1 models demonstrated performance on many benchmarks that seriously rivaled some proprietary competitors, and DeepSeek’s R1 model, released openly in January 2025, showed reasoning performance competitive with leading closed models while reportedly being trained far more cheaply, directly challenging the assumption that only the most well-funded closed labs could produce frontier-level capability. This narrowing gap suggests that whatever structural advantage proprietary labs once held through superior compute and talent concentration is not permanent or unassailable, particularly as algorithmic efficiency improvements allow smaller, more resource-constrained teams to achieve results that previously required vastly larger budgets.
The Business Case for Each Approach
From a business perspective, the two approaches reflect very different strategies for generating value. Proprietary labs like OpenAI and Anthropic generate revenue directly from API usage fees and subscription products, meaning their business model depends on keeping the model itself as the product customers pay for. Companies behind open-weight models, most notably Meta, generally do not sell model access as their primary business; Meta’s chief AI scientist Yann LeCun has argued that open models serve Meta’s interests by building goodwill, attracting developer talent, and encouraging an ecosystem of tools and applications that ultimately benefit the broader company even without directly charging for the models themselves. Mistral occupies a middle position, releasing some models openly to build community trust while keeping its most capable flagship models proprietary, reflecting a hybrid strategy that tries to capture benefits from both approaches.
Why Businesses Choose One Over the Other
For companies deciding which type of model to build on, the choice often comes down to factors beyond raw benchmark performance. Businesses with strict data privacy or regulatory requirements, such as those in healthcare, finance, or government, frequently prefer open-weight models precisely because they can run entirely within the company’s own infrastructure, with no data ever leaving their control to reach an external company’s servers. Developers building highly specialized applications may prefer to fine-tune an open-weight model specifically for their narrow use case, something that is far more difficult or impossible with a fully proprietary model where the underlying weights remain inaccessible. On the other hand, companies that prioritize simplicity, cutting-edge capability, and dedicated support, without the burden of managing their own AI infrastructure, often find a proprietary API more practical despite the recurring usage costs and the dependency on an external provider’s uptime and pricing decisions.
Security and Misuse Concerns Cut Both Ways
Each approach carries its own distinct risk profile. Proprietary models allow their creators to monitor usage, update safety measures centrally, and revoke access if a model is being misused, since every request passes through the company’s own servers. Open-weight models, once released, cannot be recalled or centrally monitored in the same way; once weights are downloaded, the releasing company has essentially no ability to prevent later misuse, modification to remove safety guardrails, or deployment in ways it never intended, a concern that has fueled ongoing debate among AI safety researchers about whether increasingly capable open-weight releases pose risks that outweigh their benefits to the broader research and developer community.
So Who Is Actually Winning?
Rather than a single winner, the proprietary versus open-weight divide looks increasingly like a durable, parallel competition in which each side pressures and improves the other. Proprietary labs continue justifying premium pricing by maintaining a capability edge at the very frontier, while open-weight releases continue forcing those same labs to compete on price and accessibility by offering an increasingly capable free or low-cost alternative. For most businesses and developers, the most practical answer to which approach is “best” depends entirely on specific priorities around cost, control, data privacy, and raw capability, rather than any universal verdict, and the healthiest sign for the industry as a whole may be that neither side has definitively defeated the other, keeping both proprietary and open-weight AI development moving forward in active competition rather than allowing either approach to become complacent.