AWS shipped a harness that costs 77% less than Claude Code on the same model
The number in the headline is the whole story, and it is a number almost nobody publishes.
AWS's Strands Agents team released Strands harness, Apache 2.0, Python and TypeScript. Not an SDK for building your own β a fully assembled, general-purpose agent you start with one line of code and run locally or drop in a Linux container. It hit the HN front page today and the harness-sdk repo is trending on GitHub alongside it.
The claim: 28% lower token cost than other harnesses running the same Claude or GPT models, across six benchmarks, at near-equal accuracy. And the sharper one: with Fable 5 on Terminal-Bench 2.1, it cost 77% less than Claude Code and scored higher. Same model underneath. All the difference is the wrapper.
The mechanics are unglamorous and that's why they work. Tool results over roughly 1,500 tokens get truncated. Context gets compacted above 85% usage. Context overflow is recovered inside the loop rather than blowing up the run. Shell, file, and web tools are built in. Model support spans Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM, so the cost comparison isn't locked to one vendor's pricing. pip install strands-harness, or npm install @strands-agents/harness.
Two labs cut prices this week and neither shipped a capability jump. This is the other lever on the same problem and the more interesting one, because it belongs to whoever writes the loop rather than whoever trains the model. Cost per completed task is the number that actually matters for agent workloads and it is the one number no model card gives you. AWS just published theirs and pointed it at a competitor's product by name.
Treat the 77% as a vendor-run benchmark until somebody reproduces it β one harness team measuring another harness on a task suite is not a neutral setup. But the framing is right, and the general-purpose positioning is a deliberate swipe: most of what people call agent harnesses are coding harnesses with the coding filed off.
https://strandsagents.com/blog/introducing-strands-harness/
← Back to all articles
AWS's Strands Agents team released Strands harness, Apache 2.0, Python and TypeScript. Not an SDK for building your own β a fully assembled, general-purpose agent you start with one line of code and run locally or drop in a Linux container. It hit the HN front page today and the harness-sdk repo is trending on GitHub alongside it.
The claim: 28% lower token cost than other harnesses running the same Claude or GPT models, across six benchmarks, at near-equal accuracy. And the sharper one: with Fable 5 on Terminal-Bench 2.1, it cost 77% less than Claude Code and scored higher. Same model underneath. All the difference is the wrapper.
The mechanics are unglamorous and that's why they work. Tool results over roughly 1,500 tokens get truncated. Context gets compacted above 85% usage. Context overflow is recovered inside the loop rather than blowing up the run. Shell, file, and web tools are built in. Model support spans Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM, so the cost comparison isn't locked to one vendor's pricing. pip install strands-harness, or npm install @strands-agents/harness.
Two labs cut prices this week and neither shipped a capability jump. This is the other lever on the same problem and the more interesting one, because it belongs to whoever writes the loop rather than whoever trains the model. Cost per completed task is the number that actually matters for agent workloads and it is the one number no model card gives you. AWS just published theirs and pointed it at a competitor's product by name.
Treat the 77% as a vendor-run benchmark until somebody reproduces it β one harness team measuring another harness on a task suite is not a neutral setup. But the framing is right, and the general-purpose positioning is a deliberate swipe: most of what people call agent harnesses are coding harnesses with the coding filed off.
https://strandsagents.com/blog/introducing-strands-harness/
Comments