Amazon distills Anthropic models to dodge token pricing hike

Key Takeaways

• Amazon engineers are distilling Anthropic's Claude models to create smaller, cheaper versions for internal use before new token-based pricing takes effect
• The partnership is being renegotiated, with Amazon shifting from compute-hour billing to per-token pricing starting next year
• Amazon is hedging by exploring alternatives including OpenAI and its own Nova models, having invested billions in both this year

Some Amazon engineers are distilling Anthropic's Claude models into smaller, cheaper alternatives ahead of a pricing overhaul that could significantly raise costs. The Information reports that the effort stems from concerns about a renegotiated partnership deal where Amazon will pay Anthropic per token processed rather than by compute hours, starting next year.

Model distillation trains a smaller model to mimic a larger one's outputs. The result is a system that approximates the original's performance at a fraction of the inference cost. Amazon has contractual rights to use Anthropic's models for this purpose, according to a source familiar with the arrangement. Apple reportedly has a similar deal with Google for Gemini.

Why is Amazon worried about token-based pricing?

Compute-hour billing lets heavy users amortize costs across large workloads. Token-based pricing, by contrast, charges for every input and output token processed. For companies running millions of queries daily, that switch can turn a predictable expense into an escalating one.

Amazon disputes the framing. A spokesperson told The Information that the expanded partnership "won't raise costs." Anthropic, for its part, argues its prices are competitive relative to model performance. Neither statement addresses whether Amazon's internal usage patterns would fare better or worse under the new structure.

The timing matters. Amazon's Bedrock platform already offers distillation tools, but Claude isn't available there. Only Amazon's Nova models and Meta's Llama are supported for that workflow. Engineers wanting distilled Claude capabilities have to build them outside the standard Bedrock pipeline.

Amazon's billion-dollar hedging strategy

The distillation push is one piece of a broader diversification effort. Amazon has reportedly invested up to $25 billion more in Anthropic this year, while also committing up to $50 billion in OpenAI. That's not a bet on one provider winning. It's a hedge against dependency on any single AI vendor.

Amazon is also exploring whether its homegrown Nova models and OpenAI's offerings could replace some Claude workloads. The company hasn't disclosed which internal applications use Anthropic models, but Alexa's AI capabilities and AWS's customer-facing features are likely candidates.

What distillation means for AI builders

If you're running AI workloads at scale, Amazon's move previews a cost optimization pattern you'll see more of. Distillation trades some capability for dramatically lower inference costs, often reducing model size by 10x or more while retaining 80-90% of the original's accuracy on narrow tasks.

The tradeoff depends on your use case. A distilled model fine-tuned for classification might match Claude's performance on that task. A general-purpose assistant handling open-ended queries will likely show the gap.

For teams building on Bedrock or similar platforms, the lesson is straightforward: lock in pricing terms before committing to a provider, and architect systems so you can swap models without rewriting your stack. Tools like n8n and Make can help abstract model calls behind configurable workflows, reducing vendor lock-in.

The partnership's uncertain future

Amazon and Anthropic haven't publicly confirmed the pricing renegotiation or distillation efforts. The Information's report relies on sources familiar with internal discussions, which means the final terms could differ from what engineers currently expect.

Still, the reported behavior tells a story. Amazon isn't just preparing for higher costs. It's building alternatives that reduce dependence on Anthropic entirely. Whether that's prudent risk management or a negotiating tactic, the message to Anthropic is clear: Amazon doesn't intend to pay whatever Claude costs.

Frequently Asked Questions

What is model distillation in AI?

Model distillation trains a smaller model to replicate the outputs of a larger one. The smaller model runs faster and costs less to operate while retaining much of the original's capability for specific tasks.

Why is Amazon distilling Anthropic's Claude models?

Amazon reportedly wants cheaper alternatives before switching to token-based pricing, which could increase costs for high-volume internal usage.

Can you distill Claude models on Amazon Bedrock?

Not currently. Bedrock's distillation service supports Amazon's Nova models and Meta's Llama, but not Claude.

How much has Amazon invested in Anthropic?

Amazon has invested up to $4 billion in Anthropic across multiple rounds, with reports of up to $25 billion more committed this year.

Source: The Decoder / Matthias Bastian