Coinbase cuts AI spend 50%, token ROI questions hit enterprise B2B

Coinbase slashed AI costs in half while usage grew, shifting to open-source models. The move exposes what sales leaders are quietly facing: token spend quintupled in H1 2026, but nobody can show the revenue lift. Even top B2B companies are now questioning whether their AI investment justifies the cost.

Coinbase cuts AI spend 50%, token ROI questions hit enterprise B2B

Coinbase cut AI spending by 50% this quarter while usage actually increased. The shift came from moving to open-source models and smarter routing, not from usage caps or friction.

That is the headline. Here is what matters for sales teams: companies quintupled their token spend in the first half of 2026. Almost nobody can point to corresponding revenue growth.

Coinbase CEO Brian Armstrong detailed the approach: better defaults, intelligent model routing, and aggressive caching. Engineers can still choose any model, but the system defaults to cheaper options unless the use case justifies frontier models. The result: flat costs, growing usage.

This hits sales organisations hard. Your sales enablement team added AI email writers, call summarisation, proposal generators, and territory planning tools. Monthly token bills went from $5k to $25k. When finance asks what changed in your pipeline metrics, you have got vague answers about productivity.

The token economics are brutal. Anthropic's Claude Sonnet 4.5 runs $3 per million input tokens, $15 per million output. The newer Fable 5 costs $10 input, $50 output. Multiply that across a 50-person sales team running AI tools daily, and the bill adds up fast. Anthropic itself generates 70-75% of its $47 billion annualized revenue from pay-per-token API calls, so they are exposed to this same question at scale.

Sales leaders need answers before Q3 planning:

What to measure: Time-to-first-meeting, proposal turnaround time, call prep efficiency, and actual close rates. If your AI spend grew 400% but your team's attainment stayed flat, you have a problem.

What to cut: Frontier models for routine tasks. An SDR personalising 50 emails does not need the most expensive model. Neither does your CRM data cleanup.

What to keep: Tools that directly impact pipeline velocity or deal size. AI that helps AEs navigate complex enterprise objections or suggests upsell opportunities can justify premium pricing.

The question is not whether AI works. The question is whether your current AI spend works. Coinbase just showed it is possible to cut costs while maintaining usage. Sales ops teams should be asking the same questions about their stack.

Anthropic filed for IPO at a $965 billion valuation, raising $65 billion in Series H. Their business model depends on enterprises continuing to burn tokens at current rates. If more companies follow Coinbase's lead, that model faces pressure.

Worth noting: this is happening at well-funded companies with sophisticated engineering teams. If Coinbase is questioning token ROI, your CFO definitely is.