A routing decision most builders made once and then left alone is starting to look like a mistake. GLM-5.2 dropped in June under an MIT license, scored 62.1% on SWE-bench Pro — above GPT-5.5 at 58.6% — and runs at $1.40 per million input tokens via Z.ai's API. Anthropic's Opus 4.8 is $5 per million input. For output tokens the gap is wider: GLM-5.2 at $4.40 versus Opus 4.8 at $25.

That's 3.6x to 5.7x cheaper, with better coding benchmark performance than GPT-5.5. Open weights, MIT license, commercially deployable.

CNBC published a July 7 investigation confirming what the OpenRouter traffic data had been signaling for months: Chinese AI models now account for between 30% and 46% of enterprise API token usage through US developer platforms. The share of OpenRouter tokens going to Chinese models has sat above 30% every single week since February 8, 2026. The average across the prior 12 months was 11%. In the first half of 2025, the figure was 4.5%.

That's not drift. That's a step function that already fired.

Source spread

• CNBC — Chinese AI models gaining ground with US companies [builder] — the July 7 primary investigation. Confirms the 30–46% OpenRouter share and the 4.5% H1 2025 baseline. Names GLM-5.2's Vercel adoption metrics directly.
• MLQ News — Chinese models surpass 30% of US OpenRouter traffic [builder] — detailed traffic breakdown; February 8 named as the inflection point.
• OpenRouter — GLM-5.2 model page [builder] — live pricing from 25 providers; baseline is $0.95–$3.00 input depending on provider.
• LLM Stats — GLM-5.2 benchmarks [builder] — SWE-bench Pro 62.1%, context window specs, sizing.
• Cryptopolitan — GLM-5.2 wins enterprise budgets [skeptic] — draws a direct line between the Fable 5 and Mythos 5 export controls in June and accelerated Chinese model adoption.

Pros & cons

What's actually changed:

• The benchmark gap is real. GLM-5.2 outscoring GPT-5.5 on SWE-bench Pro at a fraction of the input price is a legitimately different value proposition from the Chinese model releases of 2025. This isn't weight-dumping hoping nobody evaluates seriously. It passed standard benchmarks.
• The adoption curve isn't anecdotal. On Vercel alone, GLM-5.2 saw 27x daily token volume growth and 80x customer growth in its first full week after launch. That's builders choosing it for production routes, not experiments.
• MIT license means you can self-host on your own infrastructure. No API dependency, no export control exposure. That's a meaningful difference from any proprietary model at any price.
• The Fable 5 and Mythos 5 export control period in June — 13 days of global access restrictions — accelerated adoption. Builders who needed to keep agents running found Chinese alternatives. Some stayed.

What deserves actual scrutiny:

• The CNBC investigation does not break down which Chinese models specifically are in the 30–46% share. GLM-5.2 is the headline story but DeepSeek V4, Kimi K2.6, MiniMax M2.7, and LongCat-2 are all in this cohort. Quality is not uniform across the group.
• Benchmark performance is not production performance. SWE-bench Pro is a good proxy; it doesn't fully capture reasoning quality on long-horizon tasks, context utilization, or instruction-following consistency. Opus 4.8 at 69.2% is meaningfully better on that benchmark, and the gap shows on complex agent work.
• Data residency and legal exposure. Routing production enterprise traffic through Chinese-operated APIs is a compliance conversation many teams haven't had yet, even if the model weights are MIT-licensed. Self-hosting the weights is the answer; self-hosting at 753B parameters is not a weekend project.

What builders need to know

• Run a cost audit against your current routing setup. If you're sending tasks to Opus 4.8 that don't need its quality level, you're likely leaving 70–80% cost savings on the table. GLM-5.2 is the starting comparison — also check DeepSeek V4 and Kimi K2.6 for your specific task distribution.
• Self-hosting changes the compliance picture completely. Routing through Z.ai's API means Chinese infrastructure in your data path. Self-host the MIT-licensed weights to eliminate that exposure. The hardware cost for a 753B MoE model is real but for high-volume production workloads the math can close.
• Don't conflate the whole cohort. GLM-5.2 passed independent benchmarks. Not every model in the 30–46% OpenRouter pool has the same rigor. Evaluate each candidate against your actual task distribution before committing production traffic.
• Have the compliance conversation before you need to. Healthcare, finance, and anything touching personal data has data residency questions that "MIT license" doesn't answer. The license is the permission; the hosting location is the compliance question.
• Watch the threshold negotiation on White House voluntary model review standards. An announcement is expected this week. If it captures Chinese model deployments as well as US lab releases, the compliance picture changes again.

Further reading

• CNBC — Chinese AI models gaining ground with US companies as OpenAI, Anthropic costs surge — primary investigation, July 7
• OpenRouter — GLM-5.2 model page with live pricing — 25 providers, base $0.95–$3.00/M input
• LLM Stats — GLM-5.2 benchmarks and sizing — SWE-bench Pro 62.1%, context window
• MLQ News — Chinese models surpass 30% of US OpenRouter developer traffic — traffic analysis detail
• Cryptopolitan — GLM-5.2 wins enterprise budgets as export controls gate US models
• Z.ai GLM-5.2 documentation — official model docs