AWS has published a guide for connecting OpenAI Codex to GPT-5.5 on Amazon Bedrock through a customer-operated LiteLLM gateway on Amazon ECS. The design centralises model access, budgets and usage attribution while leaving Codex’s tool execution on the developer’s workstation. AWS says its live deployment passed the Responses API compatibility checks required by Codex.
The walkthrough was validated in US East (N. Virginia), us-east-1, using the gateway alias openai.gpt-5.5 mapped to bedrock_mantle/openai.gpt-5.5. Accounts need access to the selected model in the deployment region; AWS notes that availability varies by account and region. The implementation is available in the guidance-codex repository.
Codex sends task context and tool definitions through the gateway to Bedrock. When the model requests a function call, Codex executes the tool locally under its sandbox and approval settings, then sends the result in another request. The gateway governs model requests: it neither receives a general-purpose shell in the AWS account nor replaces local approvals.
LiteLLM supports user- and team-scoped keys, approved model lists, hard budgets, and request- or token-per-minute limits. Because Bedrock sees the shared ECS task role, AWS advises retaining the originating identity in scoped keys, gateway records and telemetry. Stable aliases let infrastructure teams change upstream model mappings without tying developers’ configuration to provider-specific identifiers.
The reference stack runs LiteLLM on Fargate, with RDS PostgreSQL holding state, usage and budget data. Secrets Manager and KMS protect keys, while CloudWatch supplies operational monitoring. It includes autoscaling, health checks, deployment rollback and database backups. AWS recommends HTTPS, corporate or VPN access restrictions, and private subnets for ECS tasks and the database. Load balancing, compute, RDS, WAF, logs and inference incur charges; networking may also cost extra.
Validation goes beyond a successful text response. The included probe checks Responses object structure, continuation through previous_response_id, streaming with a completed terminal response, and a forced function call with a call ID. Its continuation check asks the model to recall a unique marker from an earlier response, exposing gateways that accept the parameter without preserving state.
AWS describes the reported pass as a compatibility result, not a load test. It calls for additional testing of concurrent sessions, long-running streams, cancellation, key revocation, failure recovery and peak traffic. Customers remain responsible for gateway availability, database operations, upgrades and incident response. AWS identifies direct Bedrock access as the lower-complexity option when native identity, IAM policies and CloudTrail logs meet requirements.
Sources
Event date: 2026-09-03. Primary source date: 2026-09-03.
