一行代码让Claude推理成本降低50%

Devs just made Claude inference 50% cheaper!

And it requires just one line of code change:

• Remove → model="claude-opus-4-8" - Add → model="ship-like/claude-opus-4-8"

I verified this in my own terminal by invoking the same Anthropic model with the same prompt.

The response came back as a standard Anthropic message with similar output, usage fields, and stop reasons. This implies that the existing parsing code stays the same.

The only difference was the cost, which was 50% cheaper than standard Anthropic rates.

Under the hood, Ship does inference-time optimization, not just model selection, and they worked with me today to show how it works.

After seeing a request, it searches over possible executions, including single models, cascades, ensembles, harnesses with tools, and interventions on model activations.

It then serves the cheapest execution that still meets the reference model's quality bar.

A naive routing table isn't capable of this since it picks a model per query, and typical setups don't ensure that the cheaper model actually preserves the original's behavior.

This is not just about routing since the catch with any cheaper execution is that it might behave differently, whether that's the output format, the tool-call patterns, or the refusals.

Ship's engineering measures that its outputs stay statistically indistinguishable from the reference model on exactly these properties.

They're not token-identical, since two calls to the same model already produce different tokens, but they remain indistinguishable in capability and behavior.

So the actual product is not just cheaper inference but also a measured commitment that the cheaper execution behaves like the model the app was built against.