AWS says its open agent runs 45% cheaper than Claude Code

AWS on September 21 open-sourced Strands Harness, a ready-made general-purpose agent built on its Strands Agents SDK. Developers can run it on their own machine or deploy it to the cloud. AWS’s pitch rests on its own testing: it says the code around a model can change cost and performance a lot, even when the model stays the same.

The headline number comes from six benchmarks. AWS averaged each harness’s score across them, compared that with the average cost per task, and says Strands Harness came out 45% cheaper than Claude Code and Codex with broadly comparable accuracy. Marc Brooker, VP and distinguished engineer at AWS, walked The New Stack through how the pieces fit. The agent was due to go live later that day.

What one call assembles

Strands Agents is the open source Python SDK AWS released in May 2025. It takes what AWS calls a “model-driven approach”: the developer supplies the model, the tools and the instructions, and the model decides how to tackle the task and when to call a tool. AWS’s launch post said teams including Amazon Q Developer, AWS Glue and VPC Reachability Analyzer were already running it in production. TypeScript support entered preview in December 2025, and in February AWS set up Strands Labs as a separate home for more experimental projects.

The SDK leaves a lot of decisions to the developer. “An SDK like the Strands Harness SDK gives you the building blocks, but you still need to decide how to manage context, persist conversations, integrate tools, and guide the agent’s behavior,” Brooker says. The harness sits one layer above the SDK and makes those decisions by default. It is a preconfigured Strands Agent set up for longer-running tasks, with AWS’s choices for how the parts work together.

Out of the box you get file, shell and web tools, plus built-in handling for context, memory, persistent sessions, prompt caching and handing work off to other agents. The Strands site describes it as a fully assembled harness with “benchmarked defaults” for the system prompt, tools, memory, sessions and context management. You reach it through a CLI or import it into Python or TypeScript with a single create_harness() call. Each agent can then be tuned for its job: its instructions, its model, the tools and capabilities it can use, and whether it may delegate to another agent.

Very little of this depends on AWS. The agent loop, tools, context management, session handling and delegation are all in the open source release, and by default they run on whatever machine runs the agent.

Where the savings come from

AWS credits most of the cost result to its context-management defaults, which come down to three behaviours. Very large tool outputs are truncated before they go back to the model. When the context window fills past a set threshold, the harness compacts it, shrinking what has piled up so later calls don’t resend the whole history. And if the context overflows anyway, the agent tries to recover inside its own loop.

The six benchmarks behind the average are ALFWorld, ContextBench, GAIA, WebShop, τ³-bench and Terminal-Bench 2.1. Terminal-Bench 2.1 shows the widest gap AWS reported. Across 89 trials, Strands Harness running Fable 5 cost $56.29 against $248.05 for Claude Code, 77% less, and it scored 69.7 against 61.8. That works out to roughly 63 cents per trial against about $2.79. DeepSeek Harness was cheaper again at $40.30, but scored lower at 59.5.

All of these figures are AWS’s own, and its wider comparison already trims the headline. AWS says DeepSeek Harness ran around 14% cheaper than Strands Harness on matched runs, and once it is added to the comparison, the 45% advantage drops to 28%.

The general point, that the harness drives cost, has independent support from earlier in September. In imec’s aistack tests, swapping the harness doubled the GPU bill while the resolve rate barely moved. Claude Code, Codex and Pi all solved 44–53% of 64 SWE-Bench Pro tasks. On GLM-5.3-Flash, the imec team’s Codex sweep cost $22.8 in rented GPU time, against $45 each for Claude Code and Pi.

For Brooker, this is the reason to ship a tuned package at all. “Getting a prototype working is one step; evaluating how those choices affect performance and cost is another,” he says. “The opportunity we saw was to package that engineering into a complete, general-purpose agent.”

Bedrock by default

The one default tied to AWS infrastructure is the model call, which goes through Amazon Bedrock, AWS’s managed service for running foundation models. Brooker says one line of code switches it to Anthropic, OpenAI or Google, or to a model running locally through Ollama. On the project site, the provider is a single object passed to the agent, and the rest of the code stays the same. Changing provider doesn’t have to mean changing the model, though a different model will change how the agent behaves.

“Different models have different strengths on reasoning, tool use, and cost,” Brooker says. “What doesn’t change: context management, sessions, tools, delegation all work the same regardless of provider. No features require Bedrock.”

Every other default can be replaced too. Developers can bring their own tools and skills, connect MCP servers, change how context is handled and pick where session state is stored. In Brooker’s words, the idea is that developers “can focus on their application’s task and domain expertise, while customizing the components that need different behavior.”

From the CLI to any cloud

AWS positions Strands Harness as a general-purpose agent, though it borrows ideas from coding harnesses such as Claude Code and Codex. The difference it stresses is deployment: Strands Harness runs on whichever cloud you choose. AWS says many Claude Code and Codex users want to run the same setup in the cloud, and this is meant to answer that.

The CLI is the prototyping tool. You choose a model, add prompts, tools and other capabilities interactively, and then export the agent as Python or TypeScript code with /export. AWS also ships an Agent Skill that teaches coding agents about Strands Harness itself, including how to add MCP servers and how to generate deployment configuration for AWS, GCP, Azure, Cloudflare and Modal. The project site also lists Strands Box, an open source sandbox in preview that can run the Strands CLI, Claude Code or Codex CLI under OS-enforced containment and a “Dogwood policy”.

What’s in it for AWS?

AWS already has a place for these agents to run. Amazon Bedrock AgentCore is its managed service for deploying and operating agents, with identity and access controls, observability and the hosting infrastructure. The New Stack reads the arrangement as a clear commercial path: teams can adopt the open source harness for free, and AgentCore is the natural next step for those who later want AWS to run the infrastructure.

Brooker says AgentCore is optional and the harness can be deployed outside AWS entirely. “AgentCore is an optional hosting layer for teams that want AWS to manage the infrastructure side,” he says. “However, all deployment paths are open for the developer to choose.”

The two products are closer than separate offerings usually are. AgentCore Harness and Strands Harness were built by the same team, though they live in separate codebases, and Brooker says work on either one can feed improvements into the other.