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

OpenAI’s Codex landed on Amazon Bedrock on April 28, 2026. This was the headline announcement of “What’s Next with AWS 2026” — and it changes the enterprise coding-agent landscape. Now AWS customers can run GPT-5.5, GPT-5.4, the Codex CLI, the Codex desktop app, and the VS Code extension entirely through Bedrock APIs, with usage applicable to existing AWS commitments. Here’s how that stacks up against Codex direct API and Claude Code.

Last verified: May 4, 2026

At a glance

Option	Status	Best for	Auth model	Apply to AWS commits
Codex on Bedrock	Limited preview (Apr 28, 2026)	AWS-first enterprises	AWS IAM	Yes
Codex direct (api.openai.com)	GA	Latest features, non-AWS shops	OpenAI API key	No
Claude Code on Bedrock	GA (longer)	Enterprises wanting Opus 4.7 / Mythos	AWS IAM	Yes

Sources: openai.com/index/openai-on-aws (Apr 28, 2026), aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents (Apr 28, 2026), aboutamazon.com/news/aws/bedrock-openai-models.

Codex on Bedrock — the new enterprise default for OpenAI

The expanded AWS-OpenAI partnership announced April 28, 2026 includes three offerings, all in limited preview:

• OpenAI models on Amazon Bedrock — GPT-5.5 and GPT-5.4 accessible through standard Bedrock APIs.
• Codex on Amazon Bedrock — Codex CLI, Codex desktop app, and VS Code extension all configurable to use Bedrock as the model backend.
• Amazon Bedrock Managed Agents, powered by OpenAI — a new managed agent runtime giving enterprises a way to deploy advanced agents inside their AWS account.

Wins:

• Unified AWS controls — IAM, CloudTrail, KMS, VPC endpoints, PrivateLink.
• All customer data is processed by Amazon Bedrock (data stays in your AWS account).
• Eligible customers can apply Codex usage toward existing AWS cloud commitments.
• No new procurement or vendor contract — uses your existing AWS agreement.
• Same Codex CLI / desktop / VS Code extension developers already know.

Loses:

• Limited preview — feature parity with Codex direct will lag weeks to months.
• Limited initial regional availability (US-East/West typically first).
• Whatever rate limit profile AWS provisions during preview, not what api.openai.com gives a paying enterprise.

Best for: AWS-committed enterprises that want OpenAI models without procurement friction, regulated industries that need AWS-native security boundaries, and teams whose data residency requires Bedrock.

Codex direct (api.openai.com) — for latest features and non-AWS shops

The direct OpenAI Codex experience remains the leading edge.

Wins:

• New model snapshots and Codex features land here first.
• ChatGPT Team / Enterprise / Edu admin tooling integrates here.
• No AWS dependency — works for GCP shops, Azure shops, on-prem teams.
• Full Codex CLI feature set without Bedrock-routing limitations.

Loses:

• No AWS commit applicability — net-new spend.
• Separate vendor contract, separate SOC 2 / DPA review.
• Data leaves your AWS account if you’re AWS-hosted.

Best for: teams that want the freshest Codex features, non-AWS environments, or workflows that need ChatGPT-side admin tooling.

Claude Code on Bedrock — the incumbent enterprise pick

Claude Code on Bedrock has been the enterprise default for AWS-committed shops wanting an autonomous coding agent for months. With the May 2026 lineup of Opus 4.7 and the Mythos Preview leading SWE-Bench Pro at ~77.8% (per llm-stats.com), it remains the strongest agent on benchmarks.

Wins:

• GA on Bedrock — past the limited-preview phase Codex is in now.
• Mythos Preview leads SWE-Bench Pro (~77.8%) — strongest pure coding model in May 2026.
• Opus 4.7 strong on Terminal-Bench 2.0 long-horizon tasks.
• Claude Skills + extended thinking work natively.

Loses:

• No GPT-5.5-specific behaviors (some teams have model preferences).
• Different tool-use idioms than Codex — teams familiar with OpenAI tool-call shapes need to re-learn.

Best for: AWS enterprises that want the best benchmark performance today, teams already using Claude Skills, or anyone who has standardized on Anthropic.

Decision tree (May 2026)

Situation	Best pick
AWS-committed enterprise, want OpenAI	Codex on Bedrock (preview)
AWS-committed enterprise, want best benchmarks	Claude Code on Bedrock
Non-AWS shop using Codex	Codex direct
Need newest GPT-5.5 features ASAP	Codex direct
Need newest Claude features ASAP	Claude Direct API
Building managed agent runtime in AWS	Amazon Bedrock Managed Agents
Highest single-model SWE-Bench Pro	Claude Mythos Preview via Bedrock
Tight VS Code workflow + OpenAI standard	Codex on Bedrock with VS Code extension

What changes at the org level

Before April 28, 2026, AWS enterprises wanting OpenAI capability had to:

1. Sign a separate contract with OpenAI.
2. Move data outside AWS for inference.
3. Manage two security boundaries (AWS + OpenAI).
4. Reconcile two billing relationships.

After April 28, all four go away if you’re willing to accept limited-preview status. That removes the last meaningful procurement objection most AWS-first enterprises had to OpenAI. Expect a migration wave — and expect Anthropic to respond by deepening its own Bedrock differentiation (Mythos GA, Skills GA, agent management features).

Cost considerations

Pricing on Codex on Bedrock has not been fully published as of May 4, 2026 (limited preview). Expected:

• Per-token pricing — likely matches OpenAI direct pricing (GPT-5.5 ~$5/M input, $15/M output) plus AWS routing margin.
• AWS commit applicability — usage counts toward Enterprise Discount Programs, Private Pricing Agreements, and Savings Plans where applicable.
• No double-billing — single Bedrock invoice rather than separate OpenAI invoice.

For most AWS-committed enterprises, the savings come from commit applicability and procurement friction reduction, not raw per-token discount.

What to do this week

If you’re an AWS-committed engineering org:

1. Apply for Codex on Bedrock preview via your AWS account team. Limited preview means quotas matter.
2. Pilot one team on Codex on Bedrock + VS Code extension before standardizing. Compare to your current Claude Code on Bedrock setup head-to-head on real tickets.
3. Don’t switch wholesale until GA. Limited preview features change. Run both in parallel through Q3 2026 at minimum.
4. Watch Anthropic’s response. Expect Mythos GA, deeper Bedrock features, or pricing moves within 60 days.

Bottom line

Codex on Amazon Bedrock (April 28, 2026) is the most important enterprise coding-agent change of Q2 2026. For AWS-committed shops, Codex on Bedrock removes the last procurement objection to OpenAI. For benchmark performance today, Claude Code on Bedrock with Mythos Preview still wins. Most enterprises will run both for 60-90 days and pick by team. Expect feature parity and full GA on Codex Bedrock by late Q3 2026.

Sources: openai.com/index/openai-on-aws (April 28, 2026), aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents, aboutamazon.com/news/aws/bedrock-openai-models, theinformation.com “OpenAI Models, Codex Now Available in AWS Bedrock,” llm-stats.com SWE-Bench Pro leaderboard May 2026.