- Coinbase cut its AI spending by roughly half by routing lower-complexity workloads to open-source models, even as total token usage rose.
- Coinbase, Shopify and Ramp are building proprietary AI agents for internal workflows while maintaining Anthropic Claude contracts for high-stakes tasks.
- Ramp’s May 2026 spending data shows Anthropic’s Claude overtook OpenAI in business AI tool spend at 34.4% versus 32.3%, driven by Claude Code adoption for its reliable instruction-following.
Frontier model contracts and internal agent builds are not competing strategies, Coinbase’s financial disclosures show they run in parallel. The company roughly halved its AI spend while token volume rose, by defaulting engineers to Chinese open-weight models, Zhipu AI’s GLM 5.2 and Moonshot AI’s Kimi K2.7 Code, through an internal LLM gateway, and reserving frontier APIs for work where error cost is real. Both model makers are currently named in a congressional security probe, a detail Armstrong’s own announcement didn’t address. Shopify and Ramp are running the same play.
The Internal Agent Push
Coinbase launched its AI coding agent, Forge, to all engineers in April 2026, targeting developer productivity and internal coding workflows. Shopify has been building Sidekick, an assistant that lets merchants manage their stores through natural language. Ramp has developed its own agentic infrastructure to handle complex finance workflows autonomously. The logic in each case is the same: tailor AI to proprietary workflows, reduce external vendor dependency for routine tasks and bring long-run operational costs down.
Coinbase CEO Brian Armstrong has said he expects around 80% of workloads could eventually run on models priced at a fraction of today’s frontier rates within 12 to 18 months, though that projection depends on continued price compression. The spending split is already visible in the company’s disclosures: lower-complexity prompts go to GLM 5.2 and Kimi K2.7 Code by default; frontier models handle the work that genuinely requires them.
Anthropic’s Enduring Role
Internal agents have not displaced external model contracts. Coinbase, Shopify and Ramp all continue to use Anthropic‘s Claude for tasks where accuracy and instruction-following reliability matter most: long document processing, complex code generation and work where hallucination carries real cost.
Ramp’s spending data, published in May 2026, captures the shift at the procurement level. For the first time, Anthropic overtook OpenAI in business AI tool spend among Ramp’s customer base, Claude took 34.4% of AI spending share against OpenAI’s 32.3%. Claude Code was Anthropic’s fastest-growing product in that period. The businesses choosing Claude consistently cite the same capabilities: reliable instruction-following on long prompts, clean code output and document processing accuracy. As enterprises scale agent deployments those operational properties are proving more decisive than raw benchmark scores.
Multihoming as Policy
The pattern visible across these three firms is not “build until we no longer need to buy.” Coinbase’s spending disclosures show tiered procurement: open-source or internal models handle high-volume, lower-stakes tasks; frontier APIs cover work where error cost is high. The reliability properties that matter most at the agent layer increasingly determine where the frontier budget goes, not model selection alone.
Ramp’s procurement data points to a related dynamic at the vendor level: firms that maintain spend across multiple providers can reweight as model pricing and capability shift. Both have moved sharply, frontier inference costs have dropped roughly 95% over the past two years. Coinbase’s disclosures add a further dimension: development, integration and the ongoing work of keeping agents calibrated as underlying models change all add overhead that sits around the API call and does not appear in per-token pricing.
The total cost of ownership for enterprise AI agents runs wider than model API spend. Ramp’s procurement data and Coinbase’s spending disclosures together point to the same conclusion: the firms managing these costs most effectively have mapped which tasks justify frontier pricing and which do not, and treat that map as an operational policy, not a one-time procurement decision.



