Amazon's Harness
The AI labs cut their prices in the same week. Amazon gave away the software that runs their models, on every cloud. The brains are getting cheap, and Amazon wants to own the body.
The decision
In the same week, the two leading AI labs both cut their prices.
On September 22, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25. Anthropic says typical workloads will cost about 40% less. The next day OpenAI shipped GPT-6 Sol at $2 and $10, and GPT-6 Luna at 10 cents and 50 cents. Both cost roughly half as much as the models they replaced.
That was the headline. It was not the chess move.
The chess move came a day earlier, from a company with no frontier model anyone talks about. On September 21, AWS open-sourced Strands Harness, the software that sits between a business and the model: the loop that plans, the tools it calls, the memory it keeps, and the recovery when something breaks. It’s free under the Apache 2.0 license. It runs models from Anthropic, OpenAI and Google, open-source models through Ollama, and anything on Amazon’s own Bedrock. AWS says it runs on Azure, Google Cloud and Cloudflare too.
Then on September 22, GPT-6 Sol and Luna went live on Amazon Bedrock the same day OpenAI announced them. Until April, that couldn’t have happened. OpenAI’s models lived on Microsoft’s cloud, exclusively.
Amazon is giving away the harness and putting every lab’s model in its stable. That is the decision.
The strategic why
Start with what a harness is, because that’s where the money is moving.
A model is a brain. An agent is a brain with hands. It reads files, calls tools, writes code, checks its own work and keeps going for hours. The harness is the hands. Claude Code is Anthropic’s harness and Codex is OpenAI’s.
Until now, the labs sold the brain and the hands as a bundle, and the hands were where customers got attached. Your workflows, your permissions and your team’s habits all live in the harness.
AWS published benchmarks that go straight at that bundle. Running the same Claude model, Strands matched Claude Code’s accuracy across 6 benchmarks at roughly 40% lower cost per task. On one terminal coding benchmark it scored 8 points higher at 77% lower cost. Independent reviewers have pointed out that every test was a coding test and that some cost figures were estimates, so take the exact numbers with care. What matters is the message. Amazon is telling every enterprise buyer that it doesn’t need the lab’s hands. Amazon’s are free and cheaper, and they work with every brain.
That matters to Andy Jassy because AWS makes its money on compute, storage and the services wrapped around them. It doesn’t need to win the model race. It needs the model race to stay a race.
When models are scarce and different from each other, the lab holds the power. It sets the price, picks the cloud and owns the customer. When models are cheap and interchangeable, power moves to whoever runs the infrastructure underneath and the workflow on top. This month’s price cuts say models are heading toward the second world, and Strands is Amazon making sure they get there faster.
Jassy said it plainly in September. He now believes AWS could become a $1 trillion annual revenue business “in time.” Today it runs at about $169 billion. You don’t get from $169 billion to $1 trillion by betting on one lab. You get there by being the place where every lab’s model runs and every agent lives.
Amazon has run this play before. It never needed the best database. It needed your data to sit on AWS. Bedrock does the same thing for models: it’s a shelf, not a product. The harness is the cart that takes things off the shelf. Amazon is building the whole store.
The price war helps. Every time OpenAI halves a price or Anthropic cuts 40%, the labs are fighting each other for workloads. Amazon collects rent on the compute no matter which lab wins. Cheaper tokens mean more tokens, and more tokens mean more AWS.
The cheaper the brain gets, the more the body is worth.
What it defends against
There are 3 threats, listed in order of how much they should worry Jassy.
The labs becoming the platform. This is the big one. If Claude Code or Codex becomes the place enterprises live, the cloud underneath turns into a utility the lab picks for them. Amazon has been closing that door from several sides at once. Its $50 billion OpenAI commitment in April came with exclusive rights to Frontier, OpenAI’s agent-building tool, plus co-development of agent runtime technology on Bedrock. Strands now gives away a portable alternative to the labs’ own harnesses. Amazon wants the agent layer neutral, and a neutral layer favors the biggest cloud.
Depending on one lab. Amazon has put billions into Anthropic, and Claude has been Bedrock’s marquee model, so Amazon’s AI story was partly Anthropic’s story. That changed in April. OpenAI renegotiated with Microsoft to end Azure exclusivity, and Amazon committed up to $50 billion to OpenAI. Having GPT-6 on Bedrock on launch day is the receipt. Amazon now hosts both leading labs, and its harness treats them as interchangeable parts. It’s no longer betting on a single horse.
Google’s migration offensive. Four days after Strands shipped, Google released an open-source agent that converts AWS Kubernetes setups to Google Cloud and opens a pull request for review. Google is going directly after AWS switching costs. AWS holds about 28% of cloud infrastructure, against 20% for Azure and 15% for Google, and Google is the one gaining share. Amazon isn’t answering with walls. It’s answering with gravity: be the place where the models, the tools and the capacity already are.
The risk
Strands is portable. That’s the pitch, and it’s also the problem.
A harness that runs on Azure and Google Cloud lowers the cost of leaving AWS too. Amazon is betting that gravity beats walls. Once the models, the data and the capacity sit on AWS, it expects customers to stay because staying is easiest, not because they’re stuck. That bet has worked for 20 years, but Amazon’s own code has never tested it this directly.
Second, the labs will fight back. They can tie their best features to their own harnesses, ship new capabilities to their own tools first, or price their models to reward their own stack. Opus 5.5’s fast mode is already available only through Anthropic’s API. Expect more of that.
Third, there’s capacity. Amazon says it can’t meet all demand through 2027, even with $220 billion of capex this year. A strategy built on “run everything here” breaks down when “here” is full. Every workload AWS turns away is a customer learning how easy Strands makes it to run somewhere else.
Last, open source is a gift you don’t control. Under the Apache license, Microsoft or Google can fork Strands, tune it for their own clouds and ship it as theirs. Amazon gave away the hands, and it can’t take them back.
The harness is free. The bet is not.