LLMs

Anthropic’s latest Sonnet model is now available on AWS, offering top-tier intelligence at standard Sonnet pricing. The model excels in coding tasks by navigating large codebases and completing multi-file changes with fewer corrections. It also improves agentic performance through precise tool calling and better state management across complex steps.

• New Sonnet generation balances high capability with cost and speed efficiency.
• Coding support includes robust debugging and refactoring in large codebases.
• Agent workflows see higher success rates with better error recovery and state holding.
• Available immediately on AWS for professional knowledge work and document drafting.

Anthropic's Claude Opus 4.8 is now accessible in AWS GovCloud (US), bringing its top-tier reasoning capabilities to government and enterprise workloads. The model enhances agentic coding by planning edits and maintaining context across long sessions in real repositories. It also improves autonomous task execution through better error recovery and obstacle navigation.

• Opus 4.8 is now live in GovCloud (US) for secure, compliant AI deployments.
• Coding agents plan before editing and hold context across long repository sessions.
• Autonomous tasks feature improved error recovery and pathfinding around obstacles.
• Model designed for trusted production work with deeper reasoning capabilities.

CORTEX identifies ungrounded content in Retrieval-Augmented Generation outputs by analyzing token-level discrepancies. It compares a language model's internal representations when processing text with and without retrieved documents. Tokens heavily influenced by the context are considered grounded, while those showing minimal document-induced change are flagged as potential hallucinations. This approach allows for precise localization of errors in long-form responses.

• Enables fine-grained detection of hallucinations in specific spans rather than entire outputs.
• Leverages internal LLM representations to measure document influence per token.
• Operates by contrasting model states with and without retrieved context.
• Addresses the common issue of localized hallucinations in long-form RAG answers.

SeKV addresses the linear memory growth of KV caches in long-context LLMs by organizing context into entropy-guided semantic spans. Unlike static compression or token eviction methods, this hierarchical semantic memory allows dynamic resolution adjustments during decoding. The approach aims to retain critical token-level details without the prohibitive GPU memory overhead of uncompressed caching.

• KV cache size scales linearly with sequence length, creating a major GPU memory bottleneck.
• SeKV uses entropy to group tokens into semantic spans for more efficient storage.
• Dynamic resolution adaptation recovers detail during generation, unlike pre-prefixed compression.
• Current methods like token eviction lose fidelity when relevant tokens are compressed.

AWS GovCloud (US-West) now offers OpenAI GPT-5.4 and NVIDIA Nemotron 3 Super 120B via the Kiro IDE and CLI. GPT-5.4 targets complex agentic workflows with a 272K context window and runs on Bedrock's durable execution engine. Nemotron provides an open-weight hybrid mixture-of-experts model that activates only 12B parameters for efficiency.

• GPT-5.4 supports multi-step agentic workflows with isolated queues for resilience.
• Nemotron 3 Super 120B is an open-weight model with high compute efficiency.
• Both models are accessible via Kiro IDE and CLI in AWS GovCloud (US-West).
• GPT-5.4 features a 272K context window and a 1.2x credit multiplier.