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

One row of AWS's own benchmark table does most of the arguing: 89 trials of Terminal Bench 2.1 cost $56.29 on its new agent against $248.05 on Claude Code, and the cheaper run scored higher. The agent is Strands Harness, an open source general-purpose agent that AWS described to The New Stack ahead of the release.

At a glance

• Strands Harness is a preconfigured agent sitting on top of the Strands Agents SDK, shipping with file, shell and web tools, context handling, memory, sessions and delegation to other agents.
• Averaged across six benchmarks, AWS puts the agent 45% cheaper per task than Claude Code and Codex at broadly comparable accuracy, and credits its context-management defaults for most of that.
• The tests are AWS's own, and the company names the awkward part itself: add DeepSeek Harness, which ran around 14% cheaper on matched runs, and 45% becomes 28%.

If you have not been following Strands: AWS launched Strands Agents in May 2025 as an open source Python SDK for building agents, on what it calls a model-driven approach. You supply the model, the tools and the instructions, and the model works out how to attack the task and when to reach for a tool. AWS later brought Strands to TypeScript, and in February set up Strands Labs as a separate home for the more experimental work.

Strands Harness is a Strands Agent with the decisions already made

Marc Brooker, VP and distinguished engineer at AWS, describes the harness as a layer above the existing SDK: a preconfigured Strands Agent that bundles the machinery an agent needs to survive a long-running task, with AWS's own defaults for how the pieces fit together.

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.

Out of the box you get file, shell and web tools, plus context handling, memory, persistent sessions, prompt caching and delegation. Per agent, you can rewrite the instructions, swap the model, restrict the tools, and decide whether it may hand work to another agent. A CLI lets you assemble one interactively and then export it as Python or TypeScript with /export.

Amazon Bedrock is the only default tied to AWS infrastructure

The agent loop, the tools, context management, session handling and delegation are all in the open source release, and AWS says they run on whatever machine is running the agent. The exception is the model call, which defaults to Amazon Bedrock. Brooker says that is the single AWS-specific default and that one line replaces it with Anthropic, OpenAI, Google, or a local model through Ollama.

Everything else is adjustable too: your own tools and skills, MCP servers, a different context strategy, a different place to keep session state. Changing provider does not change the underlying model by itself, Brooker notes, and different models differ on reasoning, tool use and cost. AWS also ships an Agent Skill to teach coding agents about Strands Harness, including how to generate deployment configuration for AWS, GCP, Azure, Cloudflare and Modal.

45% cheaper becomes 28% once DeepSeek Harness is in the table

For each harness, AWS averaged the score across six benchmarks — ALFWorld, ContextBench, GAIA, WebShop, τ³-bench and Terminal-Bench 2.1 — and set that against the average cost per task on the same tests. Against Claude Code and Codex specifically, it reports Strands Harness at 45% cheaper with broadly comparable accuracy.

Widen the comparison to include DeepSeek Harness and the figure falls to 28%, because DeepSeek Harness, by AWS's account, ran around 14% cheaper than Strands Harness on matched runs. The per-test numbers are sharpest on Terminal Bench 2.1: across 89 trials, Strands Harness running Fable 5 cost $56.29 versus $248.05 for Claude Code, 77% less, and scored 69.7 against 61.8. DeepSeek Harness came in cheaper again at $40.30, with a lower score of 59.5.

AWS credits three context defaults for the cost gap

Here is the mechanism in plain words. An agent works in a loop: the model reads the conversation so far, calls a tool, the tool's output is appended, and the loop repeats. Everything sitting in that context is re-read, and re-paid for, on every turn, so a single enormous tool output keeps charging rent for the rest of the run.

Strands Harness truncates particularly large tool outputs, compacts the context once the available window passes a set threshold, and tries to recover inside the agent loop when the context overflows anyway. Think of a colleague who pastes the two relevant lines from a log rather than the whole file. Same model, same task, different bill, which is AWS's argument for packaging those choices instead of leaving each team to rediscover them.

Every number here comes from AWS testing its own harness against competitors, and the write-up names one model, Fable 5, only for the Terminal Bench 2.1 run, not for the six-benchmark average. There is also a commercial shape to this: Bedrock AgentCore is the managed hosting layer, AgentCore Harness and Strands Harness were built by the same team in separate codebases, and Brooker says work on one can feed the other. In our view the most useful thing AWS did was publish the DeepSeek comparison that cuts its own headline figure nearly in half.

Whether the numbers hold outside AWS

The agent is set to go live later on Monday, which is the point at which anyone can rerun the six benchmarks with their own model and provider and see whether the cost gap survives contact with someone else's workload. AWS has not put a schedule on how improvements move between Strands Harness and AgentCore Harness, so the open question is how long the two codebases stay in step.