Amazon Web Services

re:Invent 2025Keynote recap and announcements

Wednesday, December 3, 2025Las Vegas

1Swami opened by declaring agentic AI as a fundamentally new way of using AI.
2Amazon Bedrock AgentCore was built to solve orchestration and reliability challenges for agents.
3Amazon SageMaker AI was framed as the answer to the operational overhead of managing agent infrastructure.
4Amazon Bedrock now offers Mistral 7B and other open-weight models including Qwen and Kimi.
5Amazon Nova Forge, which Garman launched the day before, lets customers build custom frontier models.
6Amazon Nova Act reached general availability for building reliable browser automation agents.
7Blue Origin's VP of Tech Transformation joined to discuss building agentic systems for space operations.

Amazon Nova 2 SonicSpeech-to-speech model for conversational AIGA
Amazon Nova 2 LiteFast, cost-effective reasoning model with extended thinkingGA
Amazon Nova 2 OmniMultimodal model for text, image, video and speechPreview
Amazon Nova ForgeProgram to build custom frontier models on NovaAnnounced
Amazon Nova ActService for building UI workflow automation agentsGA
Amazon Bedrock AgentCoreAdded quality evaluations and policy controls for agentsAnnounced
Amazon Bedrock open-weight modelsAdds 18 models including Mistral Large 3 and Ministral 3GA

1Garman highlighted how AWS customers like Pfizer chose AWS to transform their business.
2AWS Graviton is described as the chip powering the largest share of AWS Cloud compute added.
3Trainium3 was announced, designed to make AI training and inference workloads more efficient.
4AWS Transform was pitched as AI-powered code and mainframe modernization, cutting execution time up to 80%.
5Amazon S3 Vectors reached general availability, scaling to 2 billion vectors per index.
6AWS announced Lambda Durable Functions for multi-step workflows lasting seconds to a year.
7AWS introduced Database Savings Plans, a new flexible pricing model across AWS databases.

AWS Graviton5Fifth-generation Arm chip, best price-performance for broad workloadsGA
AWS Trainium3Custom AI chip built for more efficient training and inferenceAnnounced
AWS TransformAI-powered code, mainframe and Windows modernizationGA
Amazon S3 VectorsVector storage scaling to 2 billion vectors per indexGA
Lambda Durable FunctionsMulti-step serverless workflows lasting up to one yearGA
Database Savings PlansNew flexible pricing model across AWS database servicesAnnounced
EC2 X8aedz instancesMemory-optimized instances with 5 GHz speeds and 3 TiB memoryGA

Companies on stage

Summaries of official keynote videos and recaps; watch the source for exact wording. Logos via logo.dev; trademarks belong to their owners.