Codex vs Claude Code: cost, EU hosting, governance

TL;DR

For European teams, codex vs claude code is not a question about model quality. It is a question about where requests land. Claude Code documents six deployment paths including Amazon Bedrock with eight EU regions. Codex ships through ChatGPT or the OpenAI API, with EU processing available only as a project setting after sales approval.

What Codex and Claude Code actually are

Both are agentic coding tools that work in a terminal, in an IDE and in the cloud: reading a codebase, editing files, running commands, opening pull requests. The surfaces overlap almost completely, which is why feature tables stopped being useful some time in 2026. The difference sits underneath, in which infrastructure serves the model requests.

Surface	Codex	Claude Code
Terminal	Codex CLI	claude
IDE	Codex IDE extension	VS Code and JetBrains extensions
Cloud execution	Codex Cloud	Cloud sessions, optionally self-hosted
Chat surface	ChatGPT web, desktop, mobile	claude.ai, desktop, mobile
Integrations	GitHub, Slack, GitLab (Beta), Linear	GitHub, MCP servers

Anyone asking which tool can do more will get an answer that expires within a quarter. The question that holds is which one your security team can sign off on.

Where the models run

Claude Code documents six deployment paths and states the billing route, authentication method and enterprise features for each one. Codex has two official routes plus a configurable third: a ChatGPT subscription, an API key, or a custom model provider in the config file.

Route	Codex	Claude Code
First-party subscription	ChatGPT Plus, Pro, Business, Enterprise, Edu	Claude Pro, Max, Team, Enterprise
First-party API	OpenAI API key, billed at API rates	Claude Console, pay as you go
AWS	not documented	Amazon Bedrock, Claude Platform on AWS
Google Cloud	not documented	Google Cloud Agent Platform, formerly Vertex AI
Azure	via model_providers.azure in ~/.codex/config.toml	Microsoft Foundry
Gateway in front	custom provider block	ANTHROPIC_BASE_URL and per-provider variants

The Azure row is the one worth reading twice. Microsoft documents the Codex CLI against Azure OpenAI itself and names the purpose plainly: run the coding agent entirely on Azure infrastructure and keep the data inside your compliance boundary. That works. It also carries a cost we come back to under governance.

EU processing: where the two diverge

Residency is decided by the endpoint, not by the tool. We read the region list for Claude in Amazon Bedrock and the data controls for the OpenAI platform rather than any marketing page, and the result is asymmetric.

Route	EU processing	What the source says
Claude Code on Amazon Bedrock	Yes	Eight EU regions with an EU inference profile, 10 percent premium for regional over global endpoints
Claude Code on Claude for Teams or Enterprise	No	inference_geo accepts only us and global, workspace geo only us
Claude Code on Microsoft Foundry	Partly	On Azure-hosted deployments prompts and completions stay in Azure, only usage metadata egresses to Anthropic
Codex on ChatGPT Business or Enterprise	Partly	Residency covers only eligible content and supported workloads, regions and configurations
Codex on the OpenAI API	Yes, with conditions	Europe via eu.api.openai.com, new projects only, requires Zero Data Retention or a comparable amendment
Codex on Azure OpenAI	Yes	Requests go to your own Foundry resource
Codex Cloud	Not documented	Runs on OpenAI-managed infrastructure, VM-backed sandboxes on shared infrastructure

Three points deserve attention.

First, the shortest route to EU processing leaves the first-party path in both products. For Claude Code that route is Bedrock; for Codex it is Azure. Buying Claude Code through the Team plan gets you convenience and no EU inference, because the first-party API offers us as its only workspace geo.

Second, OpenAI binds EU residency to a project configuration set at creation time that cannot be changed afterwards. Eligibility runs through sales, and models released on or after 5 March 2026 carry a 10 percent uplift on residency endpoints. This is not a checkbox a developer ticks on their own.

Third, Codex Cloud makes no statement about regions in its security documentation. Claude Code can run cloud sessions on infrastructure you operate, currently a public beta for Team and Enterprise plans. Repository checkouts, build artifacts and secrets stay on your machines while the conversation still goes to api.anthropic.com. That is an honest partial answer rather than full isolation, and organizations with Zero Data Retention or HIPAA configurations cannot use it at all.

What each one costs

Subscription prices sit close together. Billing models do not. Claude Code is included in every paid Claude plan and Codex in every paid ChatGPT plan from Go upward, but underneath run two different logics: Claude Enterprise bills usage separately at API rates, while Codex prices purchased credits per million tokens across speed tiers.

Tier	Codex	Claude Code
Entry	Go 8 US dollars	Pro 17 US dollars annual, 20 monthly
Individual	Plus 20 US dollars	Max from 100 US dollars
Heavy use	Pro from 100 US dollars, tiers to 500	Max 20x
Team	Business 20 US dollars per user annual, 25 monthly	Team 20 US dollars per standard seat annual, 25 monthly
Large organization	Enterprise on request	Enterprise 20 US dollars per seat, annual only, 20-seat minimum
Usage based	API key at API rates	Bedrock, Agent Platform, Foundry, Console

On the Codex side, purchased credits are priced per million tokens: GPT-6.1 Terra costs 25 credits for input and 125 for output at Standard speed, with Fast at twice the rate and Ultrafast at six times. Capacity planning therefore involves a speed tier, not only a token count. We broke the Claude side down in Claude Code pricing for teams in the EU.

Governance, telemetry and retention

This is the sharpest difference between the two, and the one that rarely makes it into a tool evaluation: Claude Code can be enforced centrally, Codex cannot to the same degree.

Claude Code has managed settings that local configuration cannot override. A fleet that must work through Bedrock and nothing else gets that as policy:

{
  "allowedProviders": ["bedrock"],
  "env": {
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "AWS_REGION": "eu-central-1",
    "ANTHROPIC_BEDROCK_REGION_PREFIX": "eu"
  }
}

The eu prefix makes the opus alias resolve to eu.anthropic.claude-opus-5-5, which routes through the EU inference profile. A single developer needs only the direct form:

export CLAUDE_CODE_USE_BEDROCK=1
export AWS_REGION=eu-central-1
export ANTHROPIC_BEDROCK_REGION_PREFIX=eu

The Codex equivalent is structurally different. The Azure route lives in the user configuration:

model = "gpt-5-codex"
model_provider = "azure"

[model_providers.azure]
name = "Azure OpenAI"
base_url = "https://YOUR_RESOURCE.openai.azure.com/openai/v1"
env_key = "AZURE_OPENAI_API_KEY"
wire_api = "responses"

Two limits come from the vendor documentation itself. Codex ignores openai_base_url, model_provider and model_providers in a project-local .codex/config.toml and directs those keys to the user-level config, which makes the provider a property of the developer’s machine rather than of the repository. And Microsoft states that Entra ID support is currently not available for Codex, so authentication runs on an API key in an environment variable and identity-based access drops out as a control point.

Telemetry defaults are worth stating precisely. On Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry and Claude Platform on AWS, Claude Code keeps metrics, error reports and /feedback off by default. A team that configures nothing sends nothing non-essential on those routes, and DISABLE_TELEMETRY can be pushed fleet-wide through managed settings. The collector side is covered in Claude Code telemetry with OpenTelemetry.

On retention: commercial Claude accounts have a 30-day standard period, and Zero Data Retention exists for Claude Code on Claude for Enterprise but is not part of the standard Enterprise plan, being enabled per organization after an eligibility check. Local transcripts sit in ~/.claude/projects/ for 30 days and can be shortened with cleanupPeriodDays, while transcripts you actively send with /feedback are kept for five years. On the OpenAI side, API abuse-monitoring logs are retained for up to 30 days, Zero Data Retention requires prior approval, and the Work Cloud documentation states that ending a task or deleting a chat does not immediately purge every related artifact.

Which one fits

We deliberately avoid a verdict on model quality here, because we have not measured it. The decision can be made on verifiable properties instead:

1. You have residency obligations and you run on AWS. Claude Code on Bedrock with the eu profile. The route is documented, the regions are named, and Anthropic states that its own personnel have no access to the inference infrastructure. The setup is in Claude Code on Bedrock in the EU.
2. You have residency obligations and you run on Azure. Codex against Azure OpenAI is the direct route, with the Entra ID caveat. Claude Code on Foundry is the alternative when you need Claude models.
3. You need to enforce centrally which developer reaches which endpoint. Claude Code, because managed settings override local configuration.
4. Your teams already live in ChatGPT Business and EU processing is not a hard requirement. Codex is the shorter rollout, since accounts, billing and training already exist.
5. You need execution inside your own network. Claude Code with self-hosted environments, accepting that the conversation still reaches Anthropic.

For teams without a residency obligation the choice is a matter of taste and should follow a two-week parallel trial on real repositories rather than a feature table. What that trial looks like is in Claude Code workflows in Europe.

FAQ

What does Codex cost compared to Claude Code?

Entry team tiers are level: Codex Business is 20 US dollars per user per month billed annually, Claude Team the same per standard seat. The models underneath differ. Claude Enterprise is 20 US dollars per seat with a 20-seat minimum and bills usage separately at API rates. Codex prices purchased credits per million tokens, with Fast at twice the Standard rate and Ultrafast at six times.

Does Codex run in the EU?

On three routes yes, none of them the default. The OpenAI API offers Europe as a project region on eu.api.openai.com, for newly created projects only and conditional on Zero Data Retention or a comparable amendment. Azure OpenAI in Microsoft Foundry sends requests to your own resource. For ChatGPT Business and Enterprise, OpenAI’s own wording limits residency to eligible content and supported configurations. For Codex Cloud the documentation names no region at all.

Does Claude Code run in the EU?

On Amazon Bedrock yes, across eight EU regions from Frankfurt to Paris, routed through the EU inference profile with the eu. prefix. Regional endpoints cost 10 percent more than global ones. On the first-party route no: inference_geo accepts only us and global, and us is currently the only workspace geo available.

Can IT enforce which provider the tool uses?

With Claude Code yes. allowedProviders in managed settings restricts the permitted deployment paths and cannot be overridden by local configuration; the setting needs version 2.1.285 or later. With Codex the provider lives in user-level configuration, and project-local files are explicitly ignored for provider keys. Central enforcement would have to be rebuilt through device management or network policy.

How do the telemetry defaults differ?

Claude Code keeps metrics, error reports and /feedback off by default on Bedrock, Google Cloud Agent Platform, Microsoft Foundry and Claude Platform on AWS; on the first-party route metrics are on and can be disabled with DISABLE_TELEMETRY. For Codex we found no comparable per-provider table in the documentation. What is documented are the Work Cloud admin controls and the 30-day availability of compliance log records.

Codex or Claude Code for a team of 30 engineers?

Without a residency requirement, the answer is wherever the team already works. With one, the cloud decides: Claude Code on Bedrock for AWS, Codex on Azure OpenAI or Claude Code on Foundry for Azure. At 30 seats both subscriptions land near 600 US dollars a month, which removes price from the argument and leaves operations and governance.

Sources

1. Claude Code Docs: Enterprise deployment overview (29 October 2026)
2. Claude Code Docs: Amazon Bedrock (29 October 2026)
3. Claude Code Docs: Data usage (29 October 2026)
4. Claude Code Docs: Self-hosted environments (29 October 2026)
5. Anthropic Docs: Claude in Amazon Bedrock (29 October 2026)
6. Anthropic Docs: Data residency (29 October 2026)
7. Claude: plans and pricing (29 October 2026)
8. OpenAI Docs: Codex pricing (29 October 2026)
9. OpenAI Docs: Data controls in the OpenAI platform (29 October 2026)
10. OpenAI Docs: Codex config reference (29 October 2026)
11. OpenAI Docs: Work Cloud security (29 October 2026)
12. Microsoft Learn: Codex with Azure OpenAI in Microsoft Foundry (29 October 2026)