# AWS Weekly Roundup: Prime Day Insights, New AI Models, and Security Updates
## Prime Day 2026: A Look at the Numbers That Mattered
Amazon’s biggest shopping event of the year just wrapped up, and the backend infrastructure held up beautifully. Over four days, the systems powering Prime Day handled an astronomical volume of activity. Amazon DynamoDB managed over 59 trillion requests, hitting a peak of 192 million requests per second while maintaining response times in single-digit milliseconds. Meanwhile, AWS Lambda executed more than 2.3 trillion function calls daily, and Amazon EBS processed over 24.8 trillion input/output operations, transferring more than an exabyte of data every single day.
What stood out most to the engineering teams, though, wasn’t just the volume — it was the preparation. Before the event even started, the team ran over 44,000 deliberate failure experiments using AWS Fault Injection Service. That’s more than six times the number they conducted the year before. The idea is simple but powerful: break things on purpose, long before real customers arrive, so that when millions of people shop at once, everything just works.
Also worth noting: AWS Graviton processors now handle 49 percent of the compute power Amazon.com uses across its EC2 instances, and ECS launched roughly 158 million Fargate tasks per day — a nearly 48 percent increase compared to last year.
—
## New AI Tools Are Changing How Teams Build and Work
### A Faster, Smarter Small Model
A new version of a popular lightweight AI model is now available on Amazon Bedrock. The updated model costs about 75 percent less than its predecessor while delivering comparable or better performance across most everyday tasks. One of the most exciting additions is adjustable effort levels, meaning developers can dial precision up or down depending on what a specific task needs rather than settling for one fixed configuration across their entire application. This opens the door to running many smaller, fast sub-tasks in parallel — things like routing requests, sorting data, or making quick edits across multiple files — without blowing through a budget.
### Real-Time Voice Agents Get a Major Upgrade
Building voice-powered AI assistants just got easier. A new speech-to-speech model from Amazon Nova improves reasoning, follows instructions more accurately, and handles tool calls with better precision — all while keeping response times low. It supports seven languages, offers controllable conversation pacing, and works with a large context window of up to 256,000 tokens. Combined with a newly generally available agent framework, developers can put together a functioning voice agent in just a few lines of code. The practical impact is significant: imagine a customer support agent that can look up an order, check return eligibility, and initiate a return all within a single natural conversation.
—
## Security and Safety Get Tighter
### GuardDuty Now Monitors Database Queries
Amazon GuardDuty has expanded its reach to cover actual database queries on Aurora PostgreSQL and RDS PostgreSQL. Previously, it focused mainly on login patterns and access anomalies. Now it analyzes the queries themselves, building behavioral baselines and using machine learning alongside threat intelligence to flag unusual data activity. This matters because once someone gains valid credentials — whether through phishing or a configuration mistake — their database queries can look completely normal. This new layer of protection catches suspicious behavior at the query level without requiring any additional infrastructure or changes to the database itself.
### Introducing an Agent Sandbox
As AI agents gain the ability to run shell commands and call external APIs, the question of control becomes critical. A new open-source tool called Strands Box gives teams fine-grained control over what agents are allowed to do. It uses OS-level isolation combined with policy rules written in a custom language, so teams can enforce things like “the agent can post to a notification channel, but no more than three times in ten minutes.” Importantly, real credentials never live inside the agent’s environment — they’re attached securely at an egress gateway, reducing the attack surface significantly.
—
## Integrations That Make Enterprise Life Easier
AWS Lambda now supports OAuth authentication for self-managed Apache Kafka event sources. Previously, Lambda’s Kafka integration only worked with SASL/PLAIN, SASL/SCRAM, and mTLS authentication methods. For teams in regulated industries, that was a dealbreaker. Now, Lambda Kafka consumers can authenticate through enterprise identity providers like Amazon Cognito or Okta, bringing consistent identity governance to event streaming. This works with self-hosted clusters as well as managed services from providers like Confluent, Aiven, and Redpanda.
—
## Coming Soon: re:Invent
Mark your calendars. AWS re:Invent returns to Las Vegas from November 30 to December 4, and reserved seating is currently open. The full session catalog is live, so if you’re planning to attend, now is the time to secure your spot in chalk talks, workshops, and hands-on builder sessions — they tend to fill up quickly.
—
## Frequently Asked Questions
**Q: What is AWS Fault Injection Service and why does it matter?**
A: AWS Fault Injection Service is a tool that lets engineering teams deliberately introduce failures into their systems in a controlled way. By simulating things like server crashes, network delays, or service outages before a real event happens, teams can identify weaknesses and build more resilient architectures. The massive increase in experiments ahead of Prime Day 2026 is a great example of proactive reliability engineering.
**Q: How does the new lightweight AI model compare to the previous version?**
A: The updated model delivers similar or improved quality while being significantly more affordable — roughly 75 percent cheaper than the prior version for most use cases. The addition of adjustable effort controls means teams can optimize for speed and cost on simpler tasks while reserving higher effort for more complex ones.
**Q: Can I build a voice agent without deep AI expertise?**
A: Yes. The combination of the new Nova speech-to-speech model and the generally available agent framework makes it possible to deploy a production-ready voice agent with just a few lines of code, even if you’re relatively new to AI development.
**Q: Is Strands Box only for large enterprises?**
A: No. Strands Box is open source under the Apache 2.0 license and available as a developer preview, meaning individual developers and small teams can experiment with it just as easily as large organizations.
**Q: What databases does GuardDuty’s query monitoring support?**
A: Currently, it supports Amazon Aurora PostgreSQL and Amazon RDS for PostgreSQL. Support for additional database engines may be added in the future.
—
## Conclusion
This week’s AWS updates showcase a clear trend: the platform is getting smarter, safer, and more accessible. From massive-scale reliability testing that keeps global shopping events running smoothly to small but powerful AI model improvements that save teams real money, there’s something here for builders at every level. The new security features address real-world threats in practical ways, and the expanded Kafka integration removes long-standing barriers for enterprise teams. As always, the best way to stay on top of what’s next is to keep an eye on official AWS announcements and plan ahead for flagship events like re:Invent.
Thank you for reading
