DeepSeek and China's Open-Source Models Close the Gap on US Frontier AI

DeepSeek's newest release, DeepSeek-V4-Pro, ranks 14th on Artificial Analysis's Intelligence Index. That places it behind offerings from OpenAI, Anthropic, and even fellow Chinese contender Moonshot, whose Kimi K3 model launched earlier this year. On pure capability, the US still leads. But capability is no longer the only variable enterprises are optimising for.

The thesis here is simple. Chinese open-source models, led by DeepSeek, are winning on cost and accessibility even as they trail on raw intelligence scores. That trade-off is proving attractive enough to reshape enterprise buying decisions and, by extension, the competitive geometry of the global AI race.

Kyle Chan of the Brookings Institution frames the dynamic bluntly: Chinese models are mostly open-source, letting users customise them and access roughly 90% of a top US model's capability at a fraction of the price. For cost-motivated enterprise buyers, that ratio is compelling. It explains why DeepSeek, founded by Liang Wenfeng in 2023, has moved from a niche player to a name that rattled markets and drew a direct response from the White House.

DeepSeek first broke into global headlines in January 2025 with the release of DeepSeek R1, a model built at a fraction of the cost of comparable US systems. President Trump called it a "wake up call," urging domestic industry to stay "laser-focused." That reaction alone tells you how seriously Washington now takes the threat of cheap, capable, open alternatives emerging from Beijing.

Pricing Tells the Real Story

Numbers matter more than rhetoric here. DeepSeek V4 Pro 0813 is priced at US$1.32 per million input tokens, scoring 57 on the Intelligence Index. That's not cheap relative to other open-weight models of similar size. Z.ai's GLM-5.3-Flash, for comparison, comes in at US$0.15 per million input tokens, nearly nine times less.

Against top US closed models, though, DeepSeek looks like a bargain. Anthropic's pricing spans a tenfold range, from US$0.77 per million tokens for Claude 4.5 Haiku up to US$7.70 per million tokens for Claude Fable 5 with fallback. The top scorer on the Intelligence Index, Claude Fable 5.1, hits 66. DeepSeek's 57 is respectable but not category-leading. The gap is roughly 14% on intelligence, but the price gap running into US models can be far wider depending on which tier you compare against.

This is the calculus enterprise buyers are running. A model that delivers 90% of the capability at a quarter or a tenth of the price is not a curiosity. It's a procurement decision. And it's one that Boston Consulting Group has already flagged as a structural divergence: the US retains its edge in frontier capability, talent, and capital deployment, while China accelerates on cost-optimised models and domestic adoption.

The competitive list itself is instructive. Anthropic occupies the top three spots on the Intelligence Index with Claude Fable 5.1, Claude Opus 5, and Claude Fable 5. Meta's Muse Spark 1.3 sits third. OpenAI's GPT-6 Astra and GPT-5.6 Sol follow, alongside xAI's Grok 4.6. Chinese models cluster from eighth place downward: Moonshot's Kimi K3, Zhipu AI's GLM-5.3, Alibaba's Qwen3.8, and DeepSeek's V4 Pro at 14th. The ordering confirms the current state of play. The US front-runners hold the intelligence crown. China is building depth and breadth beneath them, at scale, and cheaply.

Both sides are chasing the same theoretical endpoint. Liang has said DeepSeek's goal is AGI, telling 36Kr in 2024 that the firm needs "new model architectures to achieve stronger model capabilities with limited resources." OpenAI's Greg Brockman has claimed the firm's Astra model may already qualify, telling the Financial Times it "represents a generational leap in capability." Whether either claim survives scrutiny is a separate question. What's clear is that both camps are racing toward the same finish line with different resource constraints.

That race carries risk beyond commercial competition. MIT's Max Tegmark, writing for the Future of Life Institute, calls an AGI race "a suicide race," arguing that competitive pressure eliminates any real opportunity to solve alignment and control problems before deployment. His concern: critical infrastructure, including nuclear and financial systems, has little protection against a poorly governed system built under competitive haste. That's a geopolitical risk, not just a market one, and it sits uncomfortably alongside the commercial optimism around cheap, capable open models.

Washington's regulatory posture adds another wrinkle. The Trump administration has told AI developers it won't subject open-weight models to voluntary safety testing, according to two sources who spoke with Reuters. Only closed models, those that don't publish underlying code, face that review. It's a distinction that inadvertently favours the exact open-source strategy Chinese labs are pursuing, even as officials frame China's open-source push as a geopolitical threat.

Accusations of intellectual property theft are also mounting. Anthropic accused DeepSeek, Moonshot, and MiniMax in February 2026 of running "industrial-scale campaigns" to "illicitly extract Claude's capabilities." Moonshot, for its part, claims Kimi K3 is the first open model to reach 2.8 trillion parameters, versus an estimated 1.5 to 2 trillion for Claude Opus 4.8, according to Financial Times sourcing. Scale claims like these are difficult to verify independently, but they underscore how aggressively Chinese labs are positioning against US incumbents.

Capital is following the momentum. DeepSeek is reportedly raising roughly 50 billion yuan, about US$7.4 billion, in its first funding round, with Tencent Holdings and CATL among the investors, per Reuters. Liang also funds the firm through his own hedge fund, High-Flyer. That's serious institutional backing for a company founded just three years ago.

What Investors Should Watch

The open-source dynamic is not confined to Chinese firms. Meta, Google, and NVIDIA all maintain open-source model programmes, meaning the pricing pressure DeepSeek exerts is part of a broader industry trend rather than a uniquely Chinese phenomenon. Investors should watch three things: whether enterprise adoption data confirms the cost-driven migration Chan describes, whether US regulatory treatment of open versus closed models shifts in response to competitive pressure, and whether intellectual property disputes between Anthropic and Chinese labs escalate into something with legal or trade consequences. The intelligence gap between DeepSeek and the frontier leaders is narrowing, but slowly. The price gap is not narrowing at all. That asymmetry is the story to track.