Coinbase has cut its internal AI spending nearly in half by defaulting its engineers to two Chinese-origin open-weight models instead of expensive U.S. frontier systems. CEO Brian Armstrong revealed the shift on Saturday, routing routine coding tasks to Zhipu AI's GLM 5.2 and Moonshot AI's Kimi K2.7 Code through an internal gateway, while reserving pricier models from Anthropic and OpenAI for complex reasoning work. The move reflects a widening cost crisis across enterprise AI as companies grapple with bills that far exceeded pre-agentic budgets.

The announcement surfaces a choice now facing every engineering organization: the cheapest high-performance AI available comes from Chinese labs operating under legal frameworks that require cooperation with Beijing's intelligence services on demand. Coinbase addresses the data-routing risk by self-hosting the open-weight models on its own servers, meaning code and queries never travel to Chinese API endpoints. That approach, however, does not eliminate all exposure concerns.

How Did Coinbase Actually Cut Its AI Costs in Half?

Armstrong outlined a three-part methodology that any organization can replicate without restricting engineer access. The savings came not from a single lever but from a combination of strategic choices:

• Model Selection: Coinbase changed which models load by default in its internal gateway, routing routine tasks like code reviews, summarization, and drafting to GLM 5.2 and Kimi K2.7 Code instead of frontier models. Engineers retain freedom to escalate to more capable systems for complex planning, but 91% were never hitting their previous usage caps anyway.
• Intelligent Routing: Coinbase's internal tooling preprocesses each prompt before dispatching it, matching task type, cache status, and per-token pricing to select the most cost-effective model capable of completing the job. A frontier model may be warranted for architectural planning; it is overkill for execution-stage work.
• Cache Optimization: After optimizing caching in LibreChat, an open-source AI platform Coinbase uses internally, the cache hit rate jumped from 5% to 60%, a 12-fold improvement that means the majority of AI queries return stored results at near-zero cost rather than triggering a fresh, billable model call.

The most significant insight Armstrong did not make explicit is this: the 12-fold cache hit rate improvement matters more than the model switch itself. When a query is warm and already cached, the underlying model's capability level becomes almost irrelevant to cost. Coinbase's infrastructure savings are primarily a workflow architecture story, not a model quality story.

Why Are Chinese Models So Much Cheaper?

The economics that make GLM 5.2 and Kimi K2.7 Code credible enterprise defaults rest on a specific architectural choice: both are Mixture-of-Experts (MoE) systems. In a MoE model, only a fraction of the total parameter count activates for each token. GLM 5.2 has roughly 744 billion total parameters but activates only about 40 billion per forward pass, while Kimi K2.7 Code activates roughly 32 billion of its one trillion parameters per token. Inference cost scales with active parameters, not total parameters.

That architectural efficiency translates directly to price. GLM 5.2 costs $1.40 per million input tokens and $4.40 per million output tokens. Anthropic's Opus 4.8 lists at $5 per million input and $25 per million output, making the GLM option roughly three to four times cheaper on inputs and nearly six times cheaper on outputs at API pricing. For an organization running millions of daily engineering queries, the gap is material even before caching compounds it.

GLM 5.2 was released to subscribers by Zhipu AI on June 13, 2026, under an MIT license, meaning organizations can download the weights, run them on their own servers, modify them, and pay only for compute. Kimi K2.7 Code, released June 12, 2026 by Moonshot AI, followed under a Modified MIT license with a revenue-based attribution clause for products above 100 million monthly active users or $20 million monthly revenue.

How Do These Chinese Models Actually Perform?

Independent evaluations confirm that GLM 5.2 is the strongest open-weight model available as of June 2026 on specific coding tasks. On SWE-bench Pro, a long-horizon coding benchmark that reflects real software engineering work, GLM 5.2 scored 62.1% versus GPT-5.5's 58.6%, a meaningful gap in favor of the Chinese model. Artificial Analysis ranked it first among all open-weight models on its Intelligence Index as of mid-June.

The picture changes on different task types. GLM 5.2 trails Claude Opus 4.8 and Gemini 3.1 Pro by five to ten percentage points on Humanity's Last Exam, a scientific reasoning benchmark, and falls behind both on GPQA-Diamond, a doctoral-level science test. On Tool-Decathlon, a multi-step tool-use benchmark, it significantly underperforms both Opus 4.8 and GPT-5.5.

That task dependency is the critical caveat that Armstrong's methodology already accounts for: the three-lever framework explicitly reserves frontier models for complex planning and novel reasoning. Kimi K2.7 Code has no published independent benchmark results as of late June 2026; every performance figure Moonshot has released comes from the company's own internal benchmark suites, not from independent leaderboards.

There is also a broader benchmark reliability concern across Chinese AI models. Enterprise AI evaluation firm Kili Technology has found that production agentic AI systems show, on average, a 37% gap between published benchmark scores and real-world deployment performance. One AI researcher who tested GLM 5.2 against GPT-5.5 on debugging tasks found it "not even close" to the OpenAI model's ability to spot problems, a result that diverges sharply from the headline coding benchmark numbers.

What Are the Broader Implications for Enterprise AI?

Zhipu AI's GLM-5.2 arrival has triggered what observers are calling a "DeepSeek moment," referring to the shock wave sent through the industry when DeepSeek released capable models at a fraction of expected cost. GLM-5.2 ranks second globally on Code Arena and tops BridgeBench reasoning at 42.8, landing within a percentage point of Anthropic's Opus 4.8 on agentic benchmarks at roughly one-fifth of the cost.

Chinese labs have released more open-weight models than the rest of the world combined in 2026. Eight Chinese labs, including DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Xiaomi's Mimo, and Zhipu, have collectively released more MIT-licensed and Apache 2.0-licensed open-weight models than all non-Chinese labs combined, with six of the models now appearing on major AI capability rankings.

The cost differential raises a fundamental question for every enterprise AI buyer: not "is it good enough?" but "what reason do I have to pay more?" That question now applies to coding, reasoning, and agentic tasks. Washington's June 26 export control round is unlikely to change the trajectory. China has already trained the models, and the open-weight releases remove the dependency on U.S. cloud infrastructure entirely. A company in Southeast Asia, Europe, or the Middle East can run GLM-5.2 on local hardware with no U.S.-licensed API key and no ongoing compliance exposure.

China is simultaneously retooling its entire higher education system as a long-term signal of commitment. Beijing has eliminated 12,200 university programs, most of them in humanities, translation, and foreign languages, and launched more than 10,000 new degrees in AI, embodied intelligence, and robotics. The restructuring aligns directly with the 15th Five-Year Plan's AI-Plus initiative, treating AI engineering talent as the binding constraint on long-run capability.

For Coinbase and enterprises like it, the immediate calculus is straightforward: self-hosting open-weight Chinese models on internal servers addresses the most direct data-routing risk while delivering substantial cost savings. Whether that trade-off proves sustainable as geopolitical tensions around AI intensify remains an open question for corporate boards and compliance teams across the industry.