Coinbase has switched to cheaper Chinese AI models and cut its AI spending in half, even as the company burns through more tokens than ever. CEO Brian Armstrong says the exchange now runs on models such as GLM 5.2 and Kimi 2.7, part of a wider rush that is putting real pricing pressure on Western AI labs.
How Coinbase cut its AI costs in half
Developers at the company can still pick whatever model they prefer, but Armstrong notes that 91 percent of them never hit their old usage limits anyway, so the switch carries little practical downside for most of the team. Behind the scenes, an automatic routing system selects the best model for each request based on the task, the price, and how cacheable the request is.
Caching turned out to be the single biggest lever. Better caching alone pushed the hit rate from 5 percent to 60 percent, a dramatic improvement that slashes how often the system pays for fresh computation rather than reusing a stored result. Developers are also coached to keep context lean and start fresh sessions for new tasks, a discipline that falls under the broader idea of context engineering.
Taken together, these moves halved spending even while overall token usage kept climbing, a counterintuitive result that shows efficiency, not just cheaper models, doing much of the work. For more on this trend, see our AI coverage.
A pricing stress test for Western AI labs
Coinbase is far from alone in making this kind of move. The CEO of startup Lindy recently switched to DeepSeek v4, and Snowflake has been testing Chinese models as cheaper alternatives to OpenAI and Anthropic.
That puts pressure on Western providers at an awkward moment. Some are eyeing IPOs and need strong growth numbers to justify the enormous sums they have raised. If a wave of customers can halve their bills by routing to Chinese models, the growth assumptions baked into those valuations start to look shaky. In effect, the migration has become a live stress test for whether those assumptions hold up.
Tokenmaxxing meets accountability
Coinbase also makes each developer's AI usage visible without capping it, echoing the "tokenmaxxing" culture reported at companies like Amazon and Meta, where employees earned kudos simply for burning through tokens. But Coinbase pairs that visibility with one rule that breaks the cycle, in Armstrong's words:
- "The more you spend on AI, the more impact we expect."
That ties heavy usage back to measurable output rather than treating raw token counts as a badge of honor. It is a pointed contrast with the leaderboard-driven approach reported elsewhere in the industry, and it reframes AI spend as an investment that has to earn its return.
What it means for AI pricing
Meanwhile, a price war between OpenAI and Anthropic is reportedly brewing too. OpenAI's GPT-5.6-Sol is said to cost the same as GPT-5.5 while being more token-efficient than rival models, and the company is offering two weaker 5.6 variants at much lower prices. The combined effect of cheap Chinese models and aggressive Western pricing points toward a buyer-friendly market where compute costs keep falling. For more on the shifting economics of the sector, follow our business news.
Coinbase AI strategy FAQ
Which AI models is Coinbase using now?
Armstrong cites Chinese models including GLM 5.2 and Kimi 2.7, selected automatically per request by a routing system that weighs task, price, and caching potential.
How did Coinbase halve its AI bill while using more tokens?
The savings came from cheaper models plus efficiency gains, most notably raising the caching hit rate from 5 percent to 60 percent and enforcing leaner context use.
Why does this pressure OpenAI and Anthropic?
If large customers can cut costs sharply by adopting Chinese models, it undermines the growth narratives Western labs rely on, especially those preparing for IPOs.
