← Back to feed

The enterprise AI market has tipped. Open-source processing on OpenRouter crossed 65% of total traffic in June, up from 34% in January. The shift took five months and appears structural: Chinese-origin models now offer performance near the frontier tier at prices that make frontier pricing look like a tax on familiarity.

Citibank Research put numbers on the gap: Chinese models average $0.18 per million tokens. The average for top US frontier models is $4.00 per million. That is a 22x price differential at comparable or near-comparable task quality for a growing share of enterprise workloads.

What Changed

The immediate trigger is AI billing visibility. As more enterprises tracked AI spend by team, the pattern became undeniable: individuals spending $35,000 per month, teams exceeding usage quotas by 200%, and companies consolidating from five internal AI tools to two to contain costs.

UBS surveyed enterprise buyers actively watching AI budgets and found 60% are shifting to cheaper models or open-source Chinese alternatives. The framing inside these companies is not that frontier models failed — it is that frontier models are overkill for the 60-70% of tasks that form the bulk of daily AI usage.

The mechanism is model routing: expensive frontier models for complex planning, reasoning, and long-context work; cheaper open-source models for execution, generation, and high-volume inference. Coinbase CEO Brian Armstrong described the company’s approach in public comments: GLM 5.2 and Kimi 2.7 are now the defaults at Coinbase’s LLM gateway, with routing logic escalating to frontier models only when task difficulty warrants it.

The Models Taking Share

GLM 5.2 from Zhipu AI and Kimi 2.7 from Moonshot are the two models cited most often in enterprise routing implementations. Both are Apache-licensed or similarly permissive, meaning they can be deployed on-premises without cloud vendor lock-in. DeepSeek V4, MiniMax M3, and Qwen 3.7 round out the commonly deployed set.

The performance argument is real. GLM-5.2 scores within one percentage point of Anthropic’s Claude Opus 4.8 on agentic benchmarks. Kimi K2.7 leads the SWE-bench Pro leaderboard among open-weight models. The capability gap between Chinese open-source and US frontier, which was measured in double digits a year ago, is now measured in single digits on a growing set of tasks.

At $0.18 per million tokens versus $4, a company routing 80% of workloads to GLM 5.2 and 20% to Claude Opus 4.8 is spending materially less than a company running 100% at frontier pricing — with minimal quality degradation for the routed majority.

The Cost Ceiling

Gartner’s projection adds longer-term pressure: AI coding costs will pass the average US developer salary by 2028 at current frontier model pricing and usage growth rates. That ceiling creates a structural economic argument for model-tier management that will not go away when token prices fall, because usage grows faster than prices fall.

Reuters reported the shift based on Citibank Research data. The sourcing confirms this is not isolated behavior but a measurable category-level change: enterprise buyers are treating model selection as a cost-management variable, not a quality maximization problem.

OpenAI and Anthropic both face this headwind. Claude’s ARPU of $16.20 per monthly user remains nearly eight times OpenAI’s $2.20, but both face the same structural exposure: enterprises that have learned to route will route, and routing incentivizes open-source usage that does not generate revenue for either US lab.

Numbers

Metric	Data
OpenRouter open-source share, January 2026	34%
OpenRouter open-source share, June 2026	65%
Chinese model average cost	$0.18/M tokens
US frontier model average cost	$4.00/M tokens
Price differential	22x
Enterprises shifting to cheaper models (UBS)	60%
Users exceeding quotas by 200%+	Documented in UBS survey
Gartner: year AI coding exceeds avg developer salary	2028