AI's Great Model Flip Is Here

A seismic shift has upended the AI landscape, a dramatic turn few predicted and even fewer articulated with conviction. Data from Vercel, a leading AI hosting company, reveals an undeniable truth: open-weights AI models are now rapidly surpassing closed-weights models like OpenAI’s GPT and Anthropic’s Claude in total usage. This isn't a future projection; the flip has already happened, and it’s profound.

This trend directly challenges the reign of AI’s perceived main characters: Sam Altman of OpenAI and Dario Amodei of Anthropic. Their companies, once the undisputed architects of the AI economy, suddenly find their market definition under siege. For years, the industry danced to their tune, but the music just changed, leaving them scrambling for a new rhythm.

Powerful, affordable, and often highly customizable models drive this revolution. Many of these formidable open-weights contenders hail from China, injecting a fierce new competitive energy. DeepSeek, for instance, recently eclipsed Anthropic in total token share percentage, capturing 25.2% compared to Anthropic's 24.5% on Vercel's platform.

The development landscape is fundamentally decentralizing, moving beyond a simple duopoly. What began as a tightly controlled market is rapidly evolving into a diverse, vibrant ecosystem, thank to these accessible and potent new tools. This isn't just about cost; it's about control, customization, and a new era of innovation.

Paradoxically, the open-weights revolution hides a stark financial reality: massive usage does not equate to massive revenue. Though DeepSeek's open-weights model now commands a staggering 25.2% of total token share, eclipsing Anthropic's 24.5%, it captures a paltry 2.8% of total model spend. Anthropic, by contrast, secures a dominant 64.6% of the market's wallet.

This disparity reveals a brutal market truth: customers pay an enormous premium for even marginal gains at the frontier. Anthropic's models, used less frequently, command ~23x more spend than DeepSeek, proving that a few percentage points of state-of-the-art performance are worth billions. Compare Claude Fable 5 at $50 per million output tokens to DeepSeek V4 Flash at just $0.18 for the same volume; the price difference is stark.

Two distinct AI markets have emerged. One caters to high-volume, "good enough" tasks, where models like DeepSeek V4 Flash offer near-zero cost efficiency for the vast majority of applications. The other serves high-stakes scenarios—think high-frequency trading or critical scientific research—where the absolute best performance, however incrementally superior, justifies a multi-billion-dollar price tag. This isn't just a model flip; it's a value bifurcation.

Beyond basic economics, enterprises recoil from platform risk. Major labs like OpenAI and Anthropic, once seen as indispensable, now represent a potential threat: they could leverage client data or pivot to become direct competitors. Opting for open-weights models effectively decentralizes this power, securing a company's strategic independence.

Open models grant absolute control over data privacy, a critical factor for any sensitive industry. This granular ownership extends to deep customization: companies fine-tune models on proprietary datasets, achieving bespoke performance unavailable via generic APIs. Harvey, for example, trains models on private legal data, delivering world-class results tailored to its niche.

This embrace of open weights has birthed a vibrant ecosystem of inference providers. Dozens of firms now compete to host and serve these models, shattering the pricing stranglehold once enjoyed by the big two labs. Such fierce competition drives innovation, pushing capabilities forward while dramatically lowering costs for end-users.

AI economy isn't collapsing; it's stratifying. Expect a clear three-tier market, each serving distinct needs: mass-market commodity models, specialized enterprise solutions, and high-stakes frontier AI. This structure benefits every player, from nascent startups to established giants.

At the base, cheap, powerful open-weights models like DeepSeek V4 Flash, priced at $0.18 per million output tokens, become the bedrock for startups and general tasks. Enterprises build proprietary moats by fine-tuning these models, creating bespoke AI for internal processes without platform risk. Finally, frontier labs like OpenAI and Anthropic, though holding only 10-25% of token share, capture 60-90% of economic value by focusing on multi-billion dollar R&D for the hardest, high-value problems—think high-frequency trading.

Open source's ascendance does not signal the demise of frontier AI; it forces a healthy competition and expands the entire market. More accessible AI drives massive demand for underlying infrastructure, creating a boon for players like Nvidia. This dynamic ensures continuous innovation, pushing the boundaries of what AI can achieve across all tiers.

What are open-weights vs. closed-weights AI models?

Closed-weights models (like OpenAI's GPT series) are proprietary systems accessed via an API. Open-weights models are publicly released, allowing anyone to download, modify, and run them on their own infrastructure, offering greater control and customization.

Why are open-weights models becoming more popular?

They offer significant advantages in cost, privacy, and customization. Businesses can fine-tune them on proprietary data without sharing it and avoid dependency on a single provider, a concept known as 'platform risk'.

Is Anthropic losing to DeepSeek?

In terms of raw usage (token volume), open models like DeepSeek are surpassing Anthropic on some platforms. However, Anthropic captures vastly more revenue, as customers pay a huge premium for its frontier performance on mission-critical tasks.

What is 'platform risk' in AI?

Platform risk is the danger of building your business on another company's core technology (like a closed AI model). The platform owner could change prices, restrict access, or even become a direct competitor using insights from your data.