AWS releases Strands harness, an open-source general-purpose agent it says costs 28 percent less than Claude Code and Codex on six benchmarks

AWS released Strands harness on 21 September 2026, a preconfigured general-purpose agent for Python and TypeScript that runs on Amazon Bedrock, Anthropic, OpenAI, Google or a local Ollama model with one line of code. The vendor blog by Arron Bailiss, Tim Moreton and Albert Zhao at strandsagents.com puts it under Apache 2.0 and says it "costs 28 percent less when using the same Claude or GPT models across six benchmarks" than Claude Code, Codex and other harnesses.

What the numbers are, and what they compare

AWS averaged each harness across six benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ³-bench and Terminal Bench 2.1, and compared average cost per task at matched accuracy. AWS told The New Stack the 45 percent gap against Claude Code and Codex specifically drops to 28 percent once DeepSeek Harness is folded into the wider comparison, because DeepSeek Harness ran about 14 percent cheaper than Strands on the same runs.

On Terminal Bench 2.1 with Fable 5, AWS reports Strands harness cost $56.29 across 89 trials while Claude Code cost $248.05, a 77 percent reduction, and Strands scored 69.7 against Claude Code's 61.8. DeepSeek Harness on the same test cost $40.30 but scored 59.5.

Where the saving comes from

AWS attributes the token efficiency to three defaults built into the harness. Tool outputs above roughly 1,500 tokens are truncated. Context is summarised once the model's window fills past 85 percent. If it still overflows, the harness attempts recovery inside the agent loop rather than failing the task. Prompt caching for reused request parts is on by default.

The harness itself is not tied to AWS infrastructure. VP and distinguished engineer Marc Brooker told The New Stack that Bedrock is the default model route but "can be easily overridden to use a different model provider with one line". The agent loop, tools, context management, session handling and delegation all run on the machine that runs the agent. Only the underlying model call goes out.

What ships and where the repository is

Out of the box, Strands harness comes with shell, file and web tools, session-based memory that resumes on a session ID, and a helper agent for delegated subtasks. A separate CLI launched the same day lets developers prototype interactively and export the resulting configuration as Python or TypeScript.

The Strands Harness SDK is at github.com/strands-agents/harness-sdk, has 7,393 stars, and was created on 14 May 2025. AWS says a research paper covering the benchmark methodology is coming.

The claim is AWS's own

The 28 percent and 77 percent figures come from AWS's own testing on EC2 with a distributed benchmarking rig called Harbor. There is no independent reproduction yet, and AWS gains most from a lower number. Read the results as vendor claims and wait for the paper. DeepSeek Harness beating Strands on cost, and Strands publishing that anyway, is the honest bit.

For a team already committed to Claude Code or Codex because "switching costs are too high", the shape here is that a working agent now sits behind one function call across five model providers and swaps the model without changing anything else in the setup. If a lower token bill was the reason to stay put, the cost of testing an alternative is now a few hours of pointing at a new model.

Source

Strands Agents, Introducing Strands harness: frontier performance with 28% lower token cost, 21 September 2026. Additional reporting and Terminal Bench numbers by Paul Sawers, The New Stack, AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex.