AgentCore left preview, Nova 2 expanded Amazon's own model catalog, and Trainium3 promises to close the cost gap with Nvidia. Here's what matters about Bedrock this year.
After a year in preview, AWS announced general availability for Bedrock AgentCore, the dedicated agent runtime that introduces four independent components: AgentCore Runtime (per-session isolation with cold starts under 500ms), AgentCore Memory (managed short- and long-term memory with configurable retention policies), AgentCore Gateway (automatic translation of existing REST APIs into MCP-compatible tools), and AgentCore Identity (OAuth credential delegation so an agent can act on a user's behalf without exposing long-lived tokens). GA includes a formal 99.9% availability SLA, something missing during the preview phase and the main blocker for adoption in banking and healthcare.
Amazon introduced Nova 2, the second generation of its own models, in three sizes (Micro, Lite, Pro) with a notable leap: Nova 2 Pro adds native video generation up to 30 seconds and document understanding with complex layouts (nested tables, scanned forms) on par with Gemini 3 in AWS's internal benchmarks. The pitch remains price: Nova 2 Lite costs a fraction of the per-token cost of Claude Haiku 4.5 or GPT-5.5 Instant, aimed squarely at high-volume, low-complexity workloads like ticket classification or content moderation.
The Trainium3 chip, announced in preview with general availability promised for Q1 2027, promises a 40% improvement in performance per watt over Trainium2. For customers training or fine-tuning their own models on SageMaker, AWS presented figures (not independently verified) of up to 30% lower training cost compared to P5 instances based on Nvidia H100 GPUs.
The Bedrock Marketplace passed 200 third-party models available under a single billing and access-control plane, including open-weight variants like DeepSeek V3.2, Qwen 3, and Llama 4.2, plus models specialized in vertical domains (legal, biomedicine) from boutique providers. The core value proposition is preventing every team within a company from managing separate contracts and credentials per model provider.
Bedrock Guardrails 2.0 adds preconfigured policies aligned to specific regulatory frameworks, including a template for the European AI Act and, more relevant for the region, an initial template aligned to Peru's AI Law and Chile's proposed risk framework. This cuts down on the manual control-mapping work that previously fell entirely on each company's compliance team.
During the event, AWS highlighted cases of banks in Colombia and Mexico using AgentCore for customer service agents with controlled access to core banking systems via Identity, and a Central American insurer that cut claims processing time from days to hours by combining Nova 2 for document OCR with Claude Opus 4.5 (available on Bedrock) for coverage analysis.
AWS's strategy of offering the best third-party models (Anthropic, Meta, Mistral) alongside its own Nova family within the same control plane remains its competitive edge over Microsoft's tighter integration with OpenAI. For architects already operating on AWS, AgentCore's GA significantly weakens the case for building a custom agent runtime on Kubernetes, though dependency on AWS's IAM ecosystem remains a real switching cost worth weighing before committing.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel