Goldman Sachs Reportedly Tells Wall Street to Run Chinese Models

Goldman Sachs reportedly told Wall Street clients on July 13 to adopt DeepSeek V4, Kimi K2.6, GLM-5, and Qwen3.5, citing 80-90% of frontier capability at 70x lower cost, per CNBC.

A year ago, Chinese open-weight models were about 4.5% of US enterprise API traffic. The current figure is 30 to 46%. A bank that answers to compliance officers reportedly told regulated clients to route real work to models built in China. That's a cost gap wide enough to override the reflex to buy American frontier.

I've been skeptical of the 'Chinese models caught up' story. The benchmarks felt cherry-picked and the deployment stories were thin. A formal recommendation from Goldman to a compliance-bound institution is different evidence than a leaderboard screenshot. Goldman isn't in the hype business.

Seventy times cheaper at 80-90% capability is the number that reorganizes a budget. Most production traffic isn't frontier-hard. You don't need GPT-5.6 to classify a support ticket or pull fields off an invoice. Paying frontier prices for that work was habit and vendor gravity, not necessity.

Stop treating 'which model' as a fixed decision. Build an eval suite against your real workload, drop DeepSeek V4 and Qwen3.5 into it, and measure. If they clear your quality bar on 70% of traffic, route that share. Keep frontier for the hard 30%. That's recurring savings, not a one-time win.

None of this settles the harder question. Running an open-weight model on your own hardware is one thing. Calling a hosted Chinese API with customer data is another, and export-control and data-residency rules haven't caught up to either. Read the license and know where the weights actually run before you wire one in.