Amazon probes OpenAI for cheaper Claude spend after token-based pricing renegotiation

Amazon is already shopping for cheaper AI options to replace or complement Anthropic’s Claude models, after a renegotiated contract will move to token-based pricing that could substantially raise Amazon’s AI costs, according to The Information. The pricing change does not take effect until next year, but the fact that Amazon is exploring alternatives now is the real tell: when cost structure changes, procurement does not wait for the invoice.

This is also not happening in a vacuum. The report says Amazon is considering options including OpenAI, even though the new token-based pricing setup is slated to start next year. In other words, Amazon is trying to prevent a future bill shock by building leverage and alternatives before the new contract math takes effect.

To understand why token-based pricing matters, you have to look at what tokens represent in practice. In most AI usage, tokens track the amount of text processed, which typically scales with how long prompts are, how many inputs are sent, and how verbose the model’s responses are. A shift to token-based pricing can be more volatile than simpler pricing structures because usage patterns, product features, and even customer behavior can change month to month. For an enterprise buyer like Amazon, that means forecast risk, not just headline cost.

And enterprise buyers tend to care about forecast risk a lot, because AI spending rarely stays “experimental” for long. Once a team integrates a model into customer-facing workflows, internal tooling, or developer platforms, usage grows. Developers build on what works, products add features, and the volume of prompts creeps upward. If Amazon’s AI costs rise meaningfully under a token-based contract, procurement pressure typically spills outward into engineering roadmaps, model-selection decisions, and vendor negotiations for everyone else providing compute and AI services.

The timing is also important. The new pricing structure does not take effect until next year, but Amazon is exploring options now. That early scrambling suggests Amazon wants two things at once: continuity and optionality. Continuity means not being forced into a last-minute vendor switch that could disrupt product timelines or service reliability. Optionality means ensuring there are viable alternatives in place if the token-based economics turn unfavorable.

There is a deeper competitive angle here, too. Anthropic’s Claude models operate in a market where model performance and deployment reliability are often only half the story. Commercial terms matter as much as the benchmark charts, because AI spending is now a core part of how large platforms sell and operate. If a major buyer like Amazon is revisiting pricing assumptions, it signals that the “best model” conversation is merging with the “best deal” conversation at board level. That can influence how future contracts are written across the entire ecosystem.

For Amazon, the strategic stakes are straightforward. A contract that shifts pricing structure can turn expected costs into surprises. If token-based pricing raises costs more than anticipated, Amazon will face a choice: pass costs through to customers, absorb them to protect margins, or change how its products use AI. Each option affects competitiveness, especially in cloud and platform segments where buyers compare value beyond raw AI quality.

For peers and investors watching the AI stack, the second-order implication is that vendor diversification becomes a cost-control strategy, not just a technical preference. The report highlights a deepening pattern: major buyers are not treating model providers as interchangeable research experiments. They are treating them like dynamic, renegotiated supply chains, where pricing structures can be renegotiated and where early planning can prevent a financial reckoning later. In practice, that means boards should watch not only model capability, but also the commercial terms that determine unit economics as usage scales.