Anthropic's Claude models are moving through Vercel's AI Gateway at a per-token price that runs 4.4 times higher than the average across every other provider on the platform — and usage hasn't budged in response. According to The Decoder, developers routing traffic through Vercel keep sending a disproportionate share of requests to Anthropic despite the markup.
That's a notable data point because Vercel's AI Gateway isn't a single-vendor showcase. It sits in front of OpenAI, Google, Anthropic, xAI, Mistral, DeepSeek, and a growing list of other model providers, giving Vercel a cross-provider view of what teams actually spend per token in production, not what vendors list on their pricing pages. A 4.4x gap on that kind of aggregate is a real market signal, not a rounding artifact from one expensive workload.
The more interesting fact is what didn't happen: nobody routed around the premium. In a market where switching providers is often a one-line config change, sustained overpayment usually means the buyer has decided price isn't the variable that matters most.
Why a gateway makes this number credible
Model pricing pages are easy to compare and easy to be misled by — headline per-token rates rarely account for prompt caching, context-window overhead, retries, or the mix of input versus output tokens a given task actually burns. A gateway sees the real bill: aggregated spend across every model a team calls, priced the way it's actually metered. That's why a 4.4x figure surfacing from Vercel's own usage data carries more weight than a vendor-to-vendor price comparison would.
It also means the gap isn't explained by one team's premium-tier habit skewing an otherwise cheap average. Vercel's baseline blends cheap, fast models with expensive ones across a large developer base; Anthropic sitting at 4.4x that blended number says something about where Anthropic has chosen to compete — not at the low end.
Why developers keep paying anyway
The Decoder's framing — developers keep paying — is the part worth sitting with, since it says more about developer priorities than about Anthropic's price list. Claude has built its developer reputation largely around coding and agentic tool use, including through Claude Code and the Claude API, workloads where a wrong output tends to cost far more in debugging time than a few extra cents in tokens; in our estimation, that's the likeliest reason the premium hasn't dented demand.
What it means for teams building on these models
For anyone architecting an LLM-backed product, the practical lesson isn't "pay more for Anthropic." It's that per-token price stopped being a reliable proxy for total cost the moment agentic workflows entered the picture. A few things worth doing before the next budget review:
- Measure cost per completed task, not cost per token — a cheaper model that needs three retries can lose to a pricier one that succeeds on the first pass.
- Use a gateway or router to see blended spend across providers instead of trusting list prices, which is exactly the visibility that produced this 4.4x figure in the first place.
- Reserve premium models for the steps in a pipeline where failure is expensive — planning, code generation, multi-step agent loops — and route simpler classification or extraction steps to cheaper models.
- Revisit the split periodically; the price-to-quality gap between providers moves fast enough that a routing decision made six months ago may no longer hold.
AiiN's takeaway
The headline number here isn't really about Anthropic's pricing strategy — it's about what developers are revealing when they don't react to it. A sustained 4.4x premium that doesn't move usage is the market saying, in aggregate, that Claude's output is worth the difference for the tasks it's being asked to do. Whether that holds as competitors close the capability gap is the thing worth watching next; for now, the data suggests price competition in the LLM market is happening at the low end, while the top is still being won on quality.