Coinbase is trying to keep artificial intelligence costs under control by routing prompts to cheaper models, even as internal token usage grows at an exponential pace.
Brian Armstrong said Coinbase is working to send prompts to less expensive models where appropriate, with some use cases already keeping costs roughly flat while token usage continues to rise sharply. The Coinbase CEO also expects a major split in AI workloads over the next 12 to 18 months, with 80% of workloads moving to models that are 99% cheaper, while the remaining 20% continue to use the most advanced systems for tasks where maximum reasoning matters.
That turns model routing into a key operating tool rather than a minor infrastructure choice. Instead of sending every request to the most expensive frontier model, companies can match each task to the lowest-cost model that still performs well enough. Simple workflows, repetitive prompts, internal search, draft generation, summarization, support tasks and standard coding help can often run on cheaper systems. More complex reasoning, scientific work, high-level agent orchestration or sensitive decision layers may still justify premium models.
For Coinbase, the cost issue is especially important because AI is no longer a side experiment. The exchange has already moved toward a leaner operating model, with Coinbase saying in May that AI is changing how engineers, nontechnical teams and automated workflows operate across the company.
Cheaper AI Models Could Change The Cost Curve
The new comments add a sharper cost angle to Coinbase’s AI strategy. A company can push AI usage higher without letting token bills scale at the same rate if it can route prompts intelligently, benchmark model quality and separate routine workloads from high-value reasoning tasks.
That creates a different AI spending model from the one many companies adopted during the first phase of generative AI adoption. Early rollouts often focused on access to the strongest available models. The next phase is more likely to focus on cost control, routing layers, evaluation systems, latency, reliability and model selection.
Armstrong’s view is that demand for intelligence is close to unlimited, but the most expensive models will not need to handle most of that demand. If that prediction holds, the winning setup inside large companies may look less like one model serving everything and more like a stack of models with different price, speed and intelligence profiles.
The same logic also affects crypto infrastructure. Coinbase has been building across AI-agent payments, developer tools, compliance automation and onchain payment rails. As software agents create more transactions, support requests, coding tasks and internal workflows, token efficiency becomes part of the operating margin story.
AI Cost Control Meets The Workforce Debate
The Coinbase update lands as finance and technology firms are already rethinking staffing around AI. Major banks are preparing for AI-driven workforce cuts as JPMorgan, Citigroup and Goldman Sachs adjust hiring, automation plans and productivity targets.
Coinbase’s angle is slightly different but connected. The exchange is not only using AI to make teams smaller or faster. It is also trying to make the AI layer itself cheaper to run. That combination matters for operating leverage: more automated output, fewer manual workflows and a lower cost per AI task.
The risk is that cheaper models are not always good enough for regulated, security-sensitive or customer-impacting work. Crypto exchanges deal with fraud detection, account restrictions, compliance checks, wallet activity, engineering systems and customer assets. Bad routing can create errors, delays or security gaps if a lower-cost model is used where stronger reasoning or human review is still required.
Coinbase’s near-term challenge is to scale AI usage without allowing model costs, accuracy issues or compliance risks to outrun the savings. Armstrong’s 12-to-18-month forecast puts a clear marker on the race: most AI workloads may become much cheaper, but companies still need routing systems that know when cheap intelligence is enough and when frontier reasoning is worth paying for.
