The shift to agentic commerce
Amazon Web Services launched an agent marketplace in partnership with Anthropic to let startups sell agents to enterprise customers. This marketplace arrived alongside the release of Claude 4 models, specifically Claude Opus 4 and Claude Sonnet 4. Claude Sonnet 4 achieved a 72.7% score on the SWE-bench, while Claude Opus 4 reached 72.5% on that benchmark. Both models include extended thinking with tool use and the ability to use tools in parallel. The deployment of Claude Commerce Agents on September 2, 2026, provides a blueprint that includes a safety harness, evaluation patterns, and a Claude Code plugin to help developers build shopping and merchant agents. These agents use Python 3.11 or later and include examples for retail, travel, telecom, and ticketing. Anthropic reports that using these agents leads to carts that are 35% larger and increases purchase completion by 60%. Anthropic also provides an engineering guide that describes the architecture, latency, and cost techniques for effective commerce agents. Claude Opus 4 remains the world’s best coding model for long-running tasks.
Pricing risks for developers
The pay-per-token model creates a significant budget risk for developers experimenting with new tools on AWS. I spent $8.43 in a single day running Claude Code on Bedrock because several attempts resulted in timeout errors. Because Bedrock charges for tokens processed during every attempt, retrying a failed prompt multiplies the cost by increasing the total volume of input text sent to the model. An Anthropic direct subscription costs approximately $20 per month and avoids the 5-hour session limits found in native setups. If you want to manage everything in one AWS bill, you must weigh that predictability against per-token costs.
| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|
| Claude Opus 4 | $15 | $75 |
| Claude Sonnet 4 | $3 | $15 |
| Grok 4.6 | $2 | $6 |
The cost for Claude Opus 4 stays at $15 per million input tokens and $75 per million output tokens. Claude Sonnet 4 costs $3 per million input tokens and $15 per million output tokens. Grok 4.6 also sits on Bedrock at $2 per million input tokens and $6 per million output tokens. Developers who build applications that provide Claude local file access find that Opus 4 creates and maintains memory files to store key information. This ability helps the model extract and save facts to maintain continuity over long tasks.
Infrastructure and enterprise scale
Amazon Bedrock provides access to more than 100 foundation models from 18 providers through a single API. Organizations use Amazon Bedrock AgentCore to handle agent runtime, identity, and observability. Anthropic and Amazon signed an agreement to commit $100 billion over ten years toward AWS technologies to secure capacity for training and deploying Claude. This capacity supports the 100,000 customers who run Claude on Bedrock. Amazon Connect also expanded its reach with agentic business solutions like Connect Decisions for supply chain and Connect Talent for hiring. Amazon Quick also entered preview as a desktop AI assistant to orchestrate actions across different platforms. Amazon continues to invest billions into this partnership to meet the growing demand for frontier models. While Salesforce Agentforce provides CRM-native automation, Amazon focuses on the infrastructure layer. Google Gemini Enterprise Agent Platform builds on Vertex AI capabilities, but Amazon Bedrock remains a model-agnostic choice for many organizations that prefer AWS. The Agent Registry launched in April to help organizations catalog and search for agents across their environment. Does the complexity of setting up IAM roles and VPC isolation outweigh the convenience of consolidated billing?
