What is Codex on Amazon Bedrock?

Codex on Amazon Bedrock is OpenAI's coding agent runtime hosted inside AWS, announced May 1, 2026 at the 'What's Next with AWS' 2026 event in limited preview. It runs the same OpenAI Codex harness as the version OpenAI hosts directly, but inside your AWS VPC with AWS-native IAM, audit logs, and governance. Backend models are GPT-5.5 and GPT-5.4 (the same OpenAI on Bedrock preview). It's distinct from Bedrock Managed Agents — Codex on Bedrock is specifically the coding agent product, while Managed Agents is a more general agent runtime.

How is Codex on Bedrock different from Codex hosted directly by OpenAI?

Three structural differences. (1) Data plane: Bedrock keeps inference inside your AWS VPC with PrivateLink and CloudTrail; OpenAI direct sends traffic to OpenAI infrastructure. (2) Procurement: Bedrock spend counts toward AWS Enterprise Discount Program credits and existing AWS contracts; OpenAI direct is a separate vendor relationship. (3) Cadence: OpenAI direct gets new Codex features first (T+0); Bedrock typically lags by days to weeks. The product capability is otherwise identical — same harness, same backend models, same general workflows.

Should I use Codex on Bedrock or Claude Code in May 2026?

Three rules. (1) AWS-native enterprise that's already committed to OpenAI as the primary model → Codex on Bedrock. (2) Multi-cloud or model-agnostic team → Claude Code, which works across Bedrock Claude, Anthropic direct, and any provider via the SDK. (3) Solo developer or small team → Codex direct from OpenAI ($20-30/month subscription) or Claude Code (subscription via Claude Pro/Max). For AWS-heavy regulated industries, Codex on Bedrock is meaningfully better than Codex direct on governance. For everyone else, the direct subscriptions are simpler and cheaper.

Can I use Claude Code on AWS Bedrock?

Yes — Claude Code can be configured to use Claude models hosted on Amazon Bedrock as the inference backend, instead of going to Anthropic direct. This is the standard pattern for AWS-native enterprises that want Claude Code's harness with Bedrock's governance. As of May 2026, AWS does not offer a 'Claude Code on Bedrock' fully managed product — you self-host the Claude Code CLI and point it at Bedrock for inference. This contrasts with Codex on Bedrock, which is a fully AWS-managed runtime.

Quick Answer

Codex on Bedrock vs Codex Direct vs Claude Code (May 2026)

Published: May 6, 2026

Codex on Bedrock vs Codex Direct vs Claude Code (May 2026)

On May 1, 2026, AWS launched Codex on Amazon Bedrock — OpenAI’s coding agent runtime hosted natively inside AWS — alongside the broader OpenAI on Bedrock preview. This creates three credible options for production coding agents: Codex on Bedrock, Codex direct from OpenAI, and Claude Code (which can run on Bedrock Claude or Anthropic direct). Here’s the comparison.

Last verified: May 6, 2026

The three options at a glance

Option	Vendor stack	Backend models	Hosting	Pricing	Best for
Codex on Bedrock	OpenAI + AWS	GPT-5.5, GPT-5.4	AWS-managed	TBD (limited preview)	AWS-native OpenAI shops
Codex direct	OpenAI	GPT-5.5, GPT-5.4	OpenAI-managed	$20-30/month subscription + API	Solo devs, small teams, multi-cloud
Claude Code	Anthropic	Claude 4.7 Opus, Sonnet 4.6	Self-host CLI; backend on Anthropic or Bedrock	$20-100/month subscription + API	Claude-committed teams, multi-cloud

Decision in 30 seconds:

AWS-native enterprise on OpenAI → Codex on Bedrock.
AWS-native enterprise on Claude → Claude Code with Bedrock backend.
Solo dev or small team on OpenAI → Codex direct ($20-30/month).
Solo dev or small team on Claude → Claude Code subscription.
Multi-model team → Claude Code (most flexible) or custom harness on AgentCore.

What’s actually new in May 2026

Three coding-agent announcements landed at “What’s Next with AWS” 2026:

OpenAI models on Bedrock (limited preview) — GPT-5.5 and GPT-5.4 as foundation models in Bedrock APIs.
Codex on Amazon Bedrock (limited preview) — OpenAI’s coding agent runtime, hosted in AWS.
Bedrock Managed Agents (powered by OpenAI) — a more general managed agent runtime, also OpenAI-powered.

Codex on Bedrock is the most direct competitor to Claude Code. Managed Agents is more general-purpose. Both share the OpenAI harness — see our Bedrock Managed Agents vs AgentCore vs Claude Agent SDK comparison for the broader agent stack.

Codex on Bedrock: what it does

The same Codex coding agent OpenAI runs directly, deployed inside AWS:

File editing across multiple files.
Test execution and iteration loops.
Multi-turn coding sessions with persistent context.
Tool use (terminal, code search, web fetch).
Same harness as OpenAI-hosted Codex.

Differences from Codex direct:

Inference path: Bedrock APIs, not OpenAI APIs.
Data plane: AWS VPC with PrivateLink option.
Audit: CloudTrail logs of every model invocation.
IAM: AWS IAM roles instead of OpenAI API keys.
Pricing: TBD; expect Bedrock token rates plus a runtime fee.
Cadence: Lags Codex direct by days to weeks for new features.

Claude Code: what it does

Anthropic’s first-party coding agent. Two deployment patterns relevant to AWS users:

Pattern 1: Claude Code with Anthropic direct backend.

Self-host the Claude Code CLI.
Inference via Anthropic’s direct API.
Subscription pricing ($20-100/month for Pro / Max tiers) or pay-as-you-go API.

Pattern 2: Claude Code with Bedrock Claude backend.

Self-host the Claude Code CLI.
Configure inference to go through Amazon Bedrock Claude.
AWS-native data plane and audit.
Pay Bedrock Claude rates ($15/$75 per 1M for Opus 4.7).

The CLI is the same; the inference backend determines governance and pricing.

Side-by-side feature comparison

Capability	Codex on Bedrock	Codex direct	Claude Code
Multi-file editing	Yes	Yes	Yes
Test execution loop	Yes	Yes	Yes
Tool use (web, terminal)	Yes	Yes	Yes
Computer use (browser actions)	No (May 2026)	No (May 2026)	Yes (via Anthropic API)
Subagents / multi-agent loops	Limited	Limited	Yes (Claude Agent SDK)
MCP support	Yes	Yes	Yes
Custom skills / commands	Yes	Yes	Yes (Claude Skills)
Plan mode	No	No	Yes
Memory across sessions	Yes	Yes	Yes
Self-host / on-prem	No (Bedrock-managed)	No	Yes (CLI)
VPC-only inference	Yes	No	Yes (via Bedrock)

Claude Code has the deepest agent feature set in May 2026 — particularly subagents, plan mode, and computer use. Codex variants have caught up on basics but trail on advanced agentic patterns.

Pricing comparison

Codex on Bedrock

Pricing not publicly disclosed at limited-preview launch. Expected components:

Bedrock GPT-5.5 / 5.4 inference rates (TBD; expect parity to +30% over OpenAI direct).
Codex runtime fee (TBD, likely per-session or per-task).
AgentCore compute (the underlying runtime).
VPC endpoint costs if using PrivateLink.

Estimated total for 100M input + 50M output tokens/month: $4-8K.

Codex direct from OpenAI

ChatGPT Plus subscription: $20/month, includes Codex with usage caps.
ChatGPT Pro: $200/month, much higher Codex limits.
ChatGPT Business / Enterprise: $25-60/user/month.
Pay-as-you-go API: GPT-5.5 at $5/$30 per 1M tokens.

Estimated total for moderate usage: $20-60/month/seat.

Claude Code

Claude Pro: $20/month, includes Claude Code with usage caps.
Claude Max: $100/month, much higher limits.
Claude Team / Enterprise: $25-60/user/month.
Pay-as-you-go API on Anthropic or Bedrock: Opus 4.7 at $15/$75 per 1M tokens.

Estimated total for moderate usage: $20-100/month/seat.

For solo developers and small teams, the subscriptions on Codex direct or Claude Code are dramatically cheaper than Codex on Bedrock at preview. Bedrock starts to make sense at enterprise scale with EDP credits.

When each one wins decisively

Codex on Bedrock wins for…

AWS-native enterprises that have committed to OpenAI as the primary model.
Regulated industries needing VPC-only inference and CloudTrail audit.
Teams already using Bedrock Managed Agents for non-coding agents who want unified billing/IAM.
Organizations with $1M+ annual AWS commits where EDP discounts apply.

Codex direct wins for…

Solo developers and small teams.
Multi-cloud or non-AWS shops.
Teams that need newest Codex features immediately on launch.
Anyone who values $20-30/month subscription procurement simplicity.

Claude Code wins for…

Teams committed to Claude long-term.
Power users who need the deepest agent capabilities (subagents, plan mode, computer use).
AWS-native enterprises that prefer Claude over GPT (run Claude Code with Bedrock Claude backend).
Multi-model teams where Claude Code’s pluggable backend is a strategic asset.

Common decision pitfalls

Five mistakes teams make when picking between these:

Overpaying for governance you don’t need. Codex on Bedrock makes sense for compliance-heavy enterprises. For solo devs, the subscription on Codex direct or Claude Code is 5-20x cheaper.
Underestimating Claude Code’s capability lead. Subagents, plan mode, and computer use are not just “nice features” — for complex coding workflows, they meaningfully outperform basic Codex loops.
Locking into single-vendor harnesses too early. Codex on Bedrock locks you into OpenAI’s harness; Bedrock Managed Agents same. Claude Code’s open backend is more strategically flexible.
Ignoring rate limits on subscriptions. Codex Plus and Claude Pro both have usage caps that can be limiting for heavy users — Pro / Max tiers or pay-as-you-go are necessary for serious use.
Mixing tooling for inconsistency. Picking Codex on Bedrock for compliance-sensitive work + Claude Code for personal productivity creates two skill sets and two billing streams. Consider standardizing.

What’s coming next

Through Q3 2026, expected developments:

Codex on Bedrock GA. Pricing publication, broader region availability, additional model support.
Claude Code on Bedrock as a managed product. AWS may launch a Claude-equivalent managed coding agent, paralleling Codex on Bedrock.
Codex feature parity with Claude Code. OpenAI is reportedly working on subagent-equivalent capabilities and a plan-mode-equivalent UX.
Open-weights coding agents. Cline, Roo Code, and OpenCode running on DeepSeek V4 / Kimi K2.6 / GLM-5.1 — the cost-optimized alternative to all three of these.

Bottom line

In May 2026, Codex on Bedrock is the right pick for AWS-native enterprises committed to OpenAI; Claude Code is the right pick for everyone else who wants enterprise-grade coding agents. Codex direct ($20-30/month subscription) wins for solo developers and small teams that don’t need AWS-native governance. The capability gap between Claude Code and Codex variants in May 2026 is real and meaningful — Claude Code has subagents, plan mode, and computer use that Codex lacks. For most teams, Claude Code with the appropriate backend (Anthropic direct for simple, Bedrock Claude for AWS-native) is the strongest default in May 2026.

Sources: AWS What’s Next with AWS 2026 announcements, AWS Bedrock OpenAI page (May 2026), Anthropic Claude Code documentation (May 2026), OpenAI Codex documentation (May 2026), Stratechery Altman/Garman interview (April 2026), Futurum Group analysis (May 2026).