Anthropic Shifts Amazon Billing to Token-Based Pricing; Amazon Evaluates Switching to OpenAI and In-House Models
0xBroomberg
Anthropic renegotiated its pricing with Amazon, shifting billing from compute hours to per-token charges starting next year; the likely cost increase is pushing Amazon to evaluate OpenAI and its own Nova models as partial replacements.
What exactly changed in the billing?
Anthropic and Amazon renegotiated pricing, switching the billing basis from compute hours to token volume, effective next year.
This means → Amazon's bill will track actual model usage directly, no longer tied to underlying compute consumption.
In plain terms = Amazon used to pay for "how many machines it rented"; now it pays for "how many questions it asked and how much text it generated" — the more it uses, the more it spends.
Where does Amazon use Claude today?
Amazon runs Anthropic models across several AI products: shopping assistant Alexa for Shopping, coding tool Kiro, and workplace assistant Quick.
These span both consumer and enterprise lines, meaning token volumes are massive — and cost sensitivity after the billing switch is high.
This reflects how deeply Amazon's product stack already depends on Anthropic; switching suppliers is not a flip of a switch.
How do both sides respond to the cost concern?
Amazon's spokesperson denied costs would rise, calling the claim "not accurate."
Anthropic's spokesperson countered that "the cost of doing important work with each generation of Claude keeps falling," noting a major Opus price cut in November 2025 and steadily improving capability per dollar.
In plain terms = each side has its own ledger — Amazon says "it won't cost more," Anthropic says "unit performance is cheaper" — but whether the total bill rises depends on whether usage growth outpaces unit-price cuts.
Who else is Amazon betting on?
Amazon agreed this year to invest up to $50 billion in OpenAI; in return, OpenAI will use AWS infrastructure, AWS will sell OpenAI models, and Amazon gains rights to use OpenAI technology in its own products.
At the same time, Amazon topped up its Anthropic investment commitment by up to $25 billion.
This means → Amazon is hedging across both leading AI labs — keeping its deep Anthropic tie while opening an alternative supply line through OpenAI.
How deep does this partnership's history run?
The relationship started in 2023: Amazon invested $4 billion, in exchange for Anthropic making AWS its primary cloud provider and using Amazon's custom AI chips Trainium and Inferentia.
When AWS sells Anthropic models, Anthropic reportedly pays cloud fees to AWS and hands over half of remaining gross profit to Amazon.
In plain terms = this is not a typical investment — Amazon is simultaneously Anthropic's shareholder, cloud landlord, and profit-sharing partner, a level of lock-in far beyond a standard commercial deal.
What about the political friction?
Amazon CEO Andy Jassy raised safety concerns about Anthropic's latest models directly with White House officials.
That move is believed to have contributed to the White House restricting foreign nationals' access to Anthropic's Mythos and Fable models; Anthropic subsequently suspended the two models more broadly.
This reflects a relationship that has moved well beyond commerce — pricing restructuring plus political divergence make whether Amazon accelerates its resource shift toward OpenAI and Nova the key variable to watch.
Content is for reference only, not financial advice.