Chinese open models cut US firms' AI bills in half

The price gap is now too large for finance chiefs to ignore. Anthropic charges $50 per million output tokens for Claude Fable 5, according to Fortune. DeepSeek-V4-Pro charges about $0.87. Moonshot's Kimi K3 sits at $15, and Z.ai's GLM-5.2 at $4.40. That spread is pulling routine engineering and document work away from the top US labs and onto Chinese open-weight models, Fortune reported on 26 July.

Coinbase is the clearest case. CEO Brian Armstrong said in late June that the company nearly halved its internal AI bill by making Zhipu's GLM 5.2 and Moonshot's Kimi K2.7 Code the defaults for engineers, routed through an internal gateway that picks a model by task and price. Caching improvements lifted cache hit rates from 5% to 60%, and 91% of engineers never hit their original usage limits. The savings came from shifting routine work — code review, document summaries — off Anthropic's and OpenAI's top-tier models, not from cutting quotas, The Decoder reported.

DoorDash is running a similar playbook with a twist: cost is not the only draw. Co-founder Andy Fang said the company hands "lower-level work" to Kimi for better quality at lower cost. Its internal DashBench test found that using Kimi K2.6 as a scout paired with Claude Fable 5 as reviewer produced the best results, at $3.81 per pull request — a hybrid approach that keeps a US frontier model in the loop rather than replacing it outright.

Bloomberg has reported that Airbnb, DoorDash and Coinbase are running these Chinese models on their own servers, not through a Chinese vendor's cloud. Moonshot pushed the competitive pressure further on 16 July with Kimi K3, described as the largest open model ever released, with open weights due to follow on 27 July.

The shift has drawn political attention. The House Homeland Security and China committees have written to DoorDash asking for a list of the Chinese models it uses and details of its security testing, with a response due by 14 August. US companies face no outright ban on using Chinese models, though some government agencies have already banned DeepSeek.

The economics are blunt: a 50-to-1 price gap on output tokens is enough to move real workloads, and Coinbase has shown the savings are material, not marginal. That puts direct pressure on Anthropic and OpenAI to defend premium pricing for anything that counts as routine work — code review, summarisation — where buyers are proving they will substitute freely. The frontier labs' safer ground is complex reasoning and review tasks, exactly where DoorDash still pairs Kimi with Claude Fable 5 rather than dropping US models entirely.

The open question is regulatory, not economic. Congress's letter to DoorDash signals that self-hosting Chinese open weights won't dodge scrutiny simply because there's no data flowing to a Chinese cloud. If security reviews start shaping which open models US companies can run internally, the price advantage Moonshot, Z.ai and DeepSeek currently enjoy could get a compliance cost attached that erodes it.