What's New

GPT-5.5, GPT-5.4, and OpenAI's Codex coding agent are now generally available on Amazon Bedrock, giving developers production-ready access to OpenAI's most capable frontier models within the AWS ecosystem. GPT-5.5 excels at agentic coding, data analysis, and long-horizon autonomous tasks, while Codex integrates directly with popular IDEs and developer toolchains with inference routed through Bedrock. Pricing matches OpenAI's first-party rates with no additional fees, and usage counts toward existing AWS commitments.

How It Works

• GPT-5.5 and GPT-5.4 are called via the OpenAI Responses API using bedrock-mantle endpoints, meaning existing OpenAI SDK code can be redirected to Bedrock by simply changing the OPENAI_BASE_URL environment variable.
• Inference runs on Bedrock's next-generation inference engine, which provides each caller an isolated queue with automated capacity management, ensuring predictable performance under heavy load.
• Durable state capture is built into the inference engine: if a node fails or restarts mid-request, the call resumes from where it left off rather than restarting from scratch.
• All calls inherit AWS-native governance controls including IAM permissions, VPC and PrivateLink network isolation, KMS encryption at rest and in transit, and AWS CloudTrail audit logging.
• Prompts and responses are not used to train models and are not shared with OpenAI, enforced at the platform level.
• Codex routes all model inference through the Responses API on Bedrock and is accessible via the Codex App, Codex CLI, and IDE integrations for Visual Studio Code, JetBrains, and Xcode.
• Bedrock supports In-Region, Geographic (Geo), and Global inference routing options, allowing teams to match data residency requirements to the appropriate routing policy.
• Authentication uses a Bedrock API key set as OPENAI_API_KEY, making the transition from direct OpenAI access straightforward for existing codebases.

Why It's Important

• Enterprises that have standardized on AWS can now access OpenAI's most capable models without leaving the AWS security and compliance perimeter, eliminating the need to manage a separate OpenAI account or data processing agreement.
• Usage counts toward existing AWS Enterprise Discount Program (EDP) or committed spend, meaning organizations can consume frontier AI capacity against budgets they have already negotiated.
• The durable, fault-tolerant inference engine is critical for long-running agentic tasks—such as multi-step code generation or autonomous data analysis—where a mid-execution failure would otherwise require a full restart.
• Codex's availability on Bedrock means more than 4 million weekly Codex users in regulated industries can now route developer AI workflows through enterprise-grade AWS controls.
• The zero-markup pricing model removes a common financial disincentive for consolidating AI spend on a cloud platform, making cost comparison straightforward.
• Data residency guarantees (processing stays within the selected Bedrock Region) unblock adoption in industries such as healthcare, finance, and government where cross-border data movement is restricted.

How It's Different

• Unlike calling OpenAI directly, Bedrock wraps GPT-5.5 and GPT-5.4 with AWS-native IAM, VPC/PrivateLink, KMS, and CloudTrail controls that integrate with existing enterprise security tooling rather than requiring separate OpenAI-side configuration.
• Bedrock's isolated per-customer queue contrasts with OpenAI's shared public API, providing more predictable latency and throughput under concurrent load.
• The durable mid-request state capture is a Bedrock infrastructure feature not present in the standard OpenAI API, making it uniquely suited for long-horizon agentic workloads.
• Pricing parity with no platform surcharge differentiates Bedrock from other cloud wrappers or third-party API gateways that typically add a margin on top of model provider rates.
• Codex on Bedrock uses pay-per-token pricing with no seat licenses or per-developer commitments, contrasting with subscription-based developer tooling models common in the market.
• The OpenAI OSS models (GPT OSS 120B, GPT OSS 20B) and safety/guardrail models (GPT OSS Safeguard 120B and 20B) are also available in the Bedrock catalog, offering a full OpenAI model family including open-weight options alongside the frontier closed models.
• AWS GovCloud support (via a linked standard account EULA acceptance flow) extends availability to US government workloads, which is not a feature of the direct OpenAI API.

When to Prefer It

• Use GPT-5.5 on Bedrock when your workload involves complex, multi-step agentic tasks—such as autonomous code refactoring across large repositories or end-to-end data analysis pipelines—where fault tolerance and sustained context are critical.
• Use GPT-5.4 on Bedrock when you need strong frontier reasoning and tool use at a better price-performance ratio than GPT-5.5, particularly for professional workflows involving reliable reasoning over complex business systems.
• Choose Codex on Bedrock when your engineering organization requires AI-assisted software development (write, refactor, debug, test) but must keep all inference within AWS for compliance or data residency reasons.
• Prefer this setup when your organization has existing AWS committed spend (EDP or savings plans) and wants AI model usage to count toward those commitments rather than generating a separate billing relationship.
• Select In-Region routing when regulations require strict single-Region data processing; select Geo routing when data must stay within a geography (e.g., EU) but you want higher throughput across multiple Regions within that boundary.
• Use the GPT OSS Safeguard models (120B or 20B) on Bedrock when you need lightweight or full-scale content moderation and guardrail enforcement integrated directly into your Bedrock-hosted AI application stack.
• Adopt this path when your security team requires unified audit logging via CloudTrail and network isolation via PrivateLink for all AI API calls, which is not achievable with direct OpenAI API access.

Availability

• Status: Generally available (GA) as of June 1, 2026, for both GPT-5.5 and GPT-5.4; Codex on Bedrock is also GA with pay-per-token pricing.
• Regional availability: Specific supported regions are listed on the AWS Regions compatibility page; In-Region, Geographic (Geo), and Global routing options are available depending on data residency needs.
• GovCloud: Available in AWS GovCloud (US) via a linked standard account EULA acceptance process before enabling in the GovCloud account.
• Pricing: Matches OpenAI first-party per-token rates with no additional Bedrock surcharge; Codex is pay-per-token with no seat licenses or per-developer commitments.
• AWS commitments: Usage counts toward existing AWS Enterprise Discount Program or other committed spend agreements.
• Additional models in catalog: GPT OSS 120B, GPT OSS 20B (general-purpose open-weight), GPT OSS Safeguard 120B, and GPT OSS Safeguard 20B (safety/moderation) are also available.
• Data privacy: Prompts and responses are not used for model training and are not shared with OpenAI; this is enforced at the Bedrock platform level.
• API compatibility: Accessible via the OpenAI SDK (Python and others) by pointing OPENAI_BASE_URL to the bedrock-mantle endpoint, minimizing migration effort for existing OpenAI API users.

Related Resources

• https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html
• https://aws.amazon.com/blogs/machine-learning/openai-models-and-codex-on-amazon-bedrock-are-now-generally-available
• https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html
• https://aws.amazon.com/blogs/aws/get-started-with-openai-gpt-5-5-gpt-5-4-models-and-codex-on-amazon-bedrock