What's New

AWS has expanded Kiro's model lineup in the AWS GovCloud (US-West) Region by adding two powerful AI models: OpenAI GPT-5.4 and NVIDIA Nemotron 3 Super 120B. GPT-5.4 brings advanced reasoning, coding, and multi-step agentic workflow capabilities to government and regulated-industry developers, while Nemotron 3 Super 120B offers a compute-efficient open-weight alternative optimized for fast agentic inference. Both models are accessible immediately via the Kiro IDE and CLI after updating to the latest version.

How It Works

• GPT-5.4 on Bedrock inference engine: Runs on Amazon Bedrock's next-generation inference engine with isolated queues and durable execution, ensuring resilient, fault-tolerant workloads even under heavy agentic task loads.
• GPT-5.4 context and credits: Offers a 272K context window enabling large document analysis and long-horizon reasoning tasks, priced at a 1.2x credit multiplier reflecting its premium capability tier.
• Nemotron 3 Super 120B architecture: Uses a hybrid mixture-of-experts (MoE) design that activates only 12B of its 120B total parameters per inference pass, dramatically reducing compute cost while maintaining high model quality.
• Nemotron context and credits: Provides a 256K context window with a 32K maximum output token limit, priced at a 0.25x credit multiplier — making it highly cost-effective for high-volume agentic workloads.
• Access mechanism: Users must update their Kiro IDE or CLI to the latest version and restart it; the new models then appear in the model selector without any additional configuration.
• GovCloud data residency: All inference requests are processed within AWS GovCloud (US-West), with cross-region traffic (from US-East) encrypted via TLS 1.2+ and content stored in the region where the Kiro profile was created.
• Spec-driven development integration: Both models plug into Kiro's spec-driven workflow — converting natural language prompts into structured requirements, architectural designs, and sequenced implementation tasks executed by parallel agents.

Why It's Important

• Government and regulated workloads gain frontier AI: Federal agencies, defense contractors, and regulated enterprises operating under FedRAMP, ITAR, or EAR constraints can now access state-of-the-art reasoning and coding models without leaving the GovCloud compliance boundary.
• Agentic engineering at scale: GPT-5.4's ability to interpret context, interact with tools, operate software environments, and verify outputs across multiple steps enables genuinely autonomous software development pipelines inside secure environments.
• Cost flexibility for high-volume use: Nemotron 3 Super 120B's 0.25x credit multiplier means teams can run large numbers of agentic tasks — code generation, documentation, test writing — at a fraction of the cost of premium models.
• Data sovereignty assurance: With content collection for service improvement disabled and no cross-region data leakage, organizations with strict data handling requirements can use these models with confidence.
• Reduced vendor lock-in risk: The availability of an open-weight model (Nemotron) alongside a proprietary one (GPT-5.4) gives GovCloud customers architectural flexibility and negotiating leverage.

How It's Different

• GovCloud-hardened inference: Unlike standard commercial deployments, GPT-5.4 in GovCloud runs with isolated queues and durable execution specifically designed for resilient, compliance-sensitive workloads — not just shared public infrastructure.
• Open-weight option in a restricted environment: Nemotron 3 Super 120B is notable as an open-weight model available inside GovCloud, a rarity compared to most AI-as-a-service offerings in regulated regions that rely exclusively on proprietary models.
• MoE efficiency advantage: Nemotron's hybrid MoE architecture activating only 12B of 120B parameters is architecturally distinct from dense transformer models, delivering faster inference and lower cost without proportional quality degradation on agentic tasks.
• Kiro's spec-driven context: Unlike raw API access to these models, Kiro wraps them in a structured development workflow (specs, hooks, property-based tests), making the models more productive for software engineering than direct prompting.
• Telemetry-off by default: GovCloud Kiro disables all usage telemetry, S3 activity reporting, and content collection — a meaningful privacy differentiation from the commercial Kiro offering.
• No autonomous agent mode: GovCloud Kiro intentionally omits the fully autonomous agent feature available commercially, reflecting a deliberate security posture appropriate for sensitive environments.

When to Prefer It

• Choose GPT-5.4 when your GovCloud workload involves complex multi-step reasoning, large document analysis (contracts, regulations, technical specs), or production agentic workflows that must reliably interpret context and verify outputs across many steps.
• Choose Nemotron 3 Super 120B when you need high-throughput, cost-sensitive agentic tasks — such as bulk code generation, automated documentation, or CI/CD-integrated test writing — where the 0.25x credit multiplier provides significant budget advantages.
• Use Nemotron for open-weight compliance requirements: If your organization's policy requires or prefers open-weight models for auditability or supply-chain transparency reasons, Nemotron is the appropriate choice.
• Use GPT-5.4 for long-context document workflows: Its 272K context window makes it the better fit when entire codebases, lengthy regulatory documents, or multi-file analysis must be processed in a single inference pass.
• Use either model for GovCloud-bound software development: Any federal or regulated-industry team building applications entirely within the GovCloud boundary should prefer these models over attempting to route traffic to commercial regions.
• Use Nemotron when latency matters at scale: Its MoE architecture's fast inference makes it preferable for interactive or near-real-time agentic loops where response speed is a priority alongside cost.

Availability

• Status: Generally available (GA) as of June 30, 2026.
• Supported region: AWS GovCloud (US-West) (us-gov-west-1); GovCloud (US-East) customers are served via cross-region inference routed through US-West.
• Access method: Available through the Kiro IDE (standalone) and Kiro CLI; VS Code plugin and social/Builder ID login are not supported in GovCloud.
• GPT-5.4 pricing: 1.2x Kiro credit multiplier; 272K context window.
• Nemotron 3 Super 120B pricing: 0.25x Kiro credit multiplier; 256K context window, 32K max output tokens.
• Key limitations: Inline code suggestions, autonomous agent mode, automated model selection (Auto), VS Code plugin, social login, telemetry, and S3 activity reporting are all disabled in GovCloud Kiro.
• Default model: Claude Sonnet 4.5 remains the default foundation model in GovCloud; GPT-5.4 and Nemotron must be selected manually via the model selector.
• Prerequisites: Users must update their Kiro IDE or CLI to the latest version and restart before the new models appear.

Related Resources

• https://docs.aws.amazon.com/govcloud-us/latest/UserGuide/govcloud-kiro.html
• https://kiro.dev/