re:Invent 2025 hidden CFM announcements guide

With re:Invent 2025 behind us and over 60 launch announcements, here are five hidden gems that may have gone unnoticed but still have big cost impacts. While everyone was talking about the database savings plans and AI announcements, these quieter launches are already helping customers optimize their cloud spend in creative ways.

1. Amazon Athena Capacity Reservation Controls

Amazon Athena now gives you granular control over Data Processing Unit (DPU) usage for queries running on Capacity Reservations. This might sound technical, but the cost impact is immediate you can now configure DPU settings at the workgroup or query level to eliminate over-provisioning and control concurrency. DPU setting controls the compute power (CPU/RAM) for your queries, allowing you to balance cost and performance using Capacity Reservations.

Previously, Capacity Reservations allocated resources generously, potentially wasting capacity on small queries, while critical queries competed for resources. Now you can set explicit DPU values per query small queries use only what they need while business-critical queries get guaranteed resources for fast execution. The Athena console shows per-query DPU usage, helping you right-size your capacity needs with real data.

2. Amazon Kinesis Video Streams Warm Storage Tier

Amazon Kinesis Video Streams introduced a warm storage tier that is cost-effective of long-term video retention. The standard tier (now called “hot tier”) remains optimized for real-time access, but the new warm tier enables extended retention at reduced costs of up to 67% for KVS stored in Kinesis Video Streams, while maintaining sub-second access latency.

The warm tier seamlessly integrates with Amazon Rekognition Video and Amazon SageMaker, so you’re not sacrificing functionality for savings. Warm tier is designed for cost optimization in use cases requiring long-term storage. If you’re retaining footage or storing training data for ML models, the warm tier delivers enterprise-grade storage economics that previously required complex tiering strategies. While the warm tier offers these cost opportunities, it requires a minimum retention period of 30 days, meaning AWS charges you for at least 30 days even if you delete the data earlier. You gain flexibility to configure fragment sizes selecting smaller fragments for lower latency use cases or larger fragments to reduce ingestion costs based on specific requirements.

3. Amazon Bedrock Reinforcement Fine-Tuning

Amazon Bedrock now supports reinforcement fine-tuning, delivering 66% accuracy gains on average over base models. While this sounds like a task optimization feature, there is a considerable cost implication on inferencethe side. You can use smaller, faster, and more cost-effective model variants while maintaining high quality. Along side this, if a model is more accurate it can result in users having to ask less followup questions which will result in an overall reduction in token input/output per engagement. The on-demand inference option includes a token-based pricing model that charges based on the number of tokens processed during inference. For pricing context the entire workflow is billed at an hourly rate. You can set up inference on a custom model by creating a custom model on-demand deployment.

Supervised fine-tuning requires large labeled datasets and expensive human annotation. Reinforcement fine-tuning takes a different approach, using reward functions to evaluate responses and teaching models to understand quality without massive amounts of pre-labeled training data.

You can downsize from higher cost models to efficient variants without sacrificing output quality. If you’re running workloads on large models for routine tasks, reinforcement fine-tuning lets you create specialized smaller models that cost less per inference. Please note that this currently works with Amazon Nova 2 Lite, with additional model support coming soon.

4. Amazon S3 Tables Intelligent-Tiering

Amazon S3 Tables now offer the Intelligent-Tiering storage class, which automatically optimizes costs based on access patterns without performance impact or operational overhead. This takes the existing S3 Intelligent-Tiering concept for S3 general purpose buckets and puts it into Amazon’s managed Apache Iceberg table storage offering of Amazon S3 Tables.

Intelligent-Tiering automatically transitions data across three access tiers: Frequent Access (default), Infrequent Access (40% lower cost than Frequent Access Tier after 30 days without access), and Archive Instant Access (68% lower cost than Infrequent Access after 90 days without access).

Using S3 Intelligent-Tiering with S3 Tables offers simplified cost optimization that is designed with managed Iceberg in mind. Maintenance operations like compaction, snapshot expiration, and unreferenced file removal keep data in their existing tier and do not move your data back to higher-cost Frequent Access tier. Compaction targets only data in the Frequent Access tier, ensuring actively queried data is optimized for performance while colder data remains in lower-cost storage. You can set Intelligent-Tiering as the storage class when creating tables or as the default for all new tables in a table bucket.

5. AWS Network Firewall Flexible Cost Allocation

This announcement is not necessarily cost saving, but allocation. AWS Network Firewall now supports flexible cost allocation through Transit Gateway native attachments, solving one of the most persistent frustrations in centralized security architectures: how do you fairly distribute firewall costs to the teams actually consuming the service?

Previously, all AWS Network Firewall data processing charges landed in the firewall owner’s account typically the central security or networking team. Now you can create metering policies that automatically allocate AWS Network Firewall data processing costs based on actual usage. Route traffic from Team A through your centralized firewall? Team A’s account gets charged for their data processing. Team B sends twice as much traffic? They pay proportionally more. The firewall owner maintains centralized security controls while costs distribute fairly to the business units generating the traffic.

Eliminate custom chargeback solutions and restore accountability for security costs. When application teams see their actual security inspection costs, they are more likely to optimize their traffic patterns, implement caching to reduce redundant inspections, and make more informed architectural decisions.

The feature works at both attachment-level and individual flow-level granularity, giving you flexibility to match your organization’s specific chargeback requirements.

Important note: This only applies to standard per-GB data processing charges not Advanced Inspection Traffic Processing or Advanced Threat Protection Traffic Processing. And there are no additional charges for using flexible cost allocation beyond standard Network Firewall and Transit Gateway pricing.

Conclusion

Beyond these announcements, re:Invent 2025 delivered significant cost optimization tools including Database Savings Plans, automated Cost Optimization Hub features, and improvements to AWS Compute Optimizer. Check out the blog for Key 2025 re:Invent Launches to Transform Your FinOps Practices to learn about all CFM announcements and see where they can be used in your organization.

You can also see us discuss these announcements and many more in 2026 by watching The Keys to AWS Optimization YouTube Channel.