Claude Code vs OpenAI Codex vs Vertex AI: Zero Data Retention for HIPAA-Regulated Engineering Teams

Where Claude’s posture is strongest

Claude appears most defensible when teams:

• use the core API path covered by a direct enterprise agreement
• route regulated inference through the Messages API
• keep PHI out of excluded features and retention-heavy workflows
• confirm contract language rather than relying on product branding

This is workable, but it demands discipline.

Where Claude’s ZDR story weakens

Several exceptions matter in healthcare engineering:

• Files uploaded through the Files API are retained until deleted
• Metrics logging, if enabled, may sit outside ZDR protection
• Code execution sandboxes are not treated the same as simple real-time inference
• Batch processing is not ZDR-eligible
• Beta features should be treated as non-ZDR unless explicitly documented otherwise
• Access through AWS Bedrock or Vertex AI follows those platforms’ privacy models, not Anthropic’s direct ZDR model

That last point is easy to miss. “Using Claude” is not one compliance posture. Direct Anthropic API, Claude via Bedrock, and Claude via Vertex are three distinct routing choices with different legal and architectural implications.

Best fit for Claude Code

Claude Code makes the most sense for teams that want advanced coding assistance and are willing to engineer around a narrow approved path. It is a fit when security, platform, and legal teams can tightly define which APIs are allowed and block everything else.

It is a weaker fit for organizations that want simple defaults. Too much depends on exact endpoint choice, feature choice, and contract scope.

What OpenAI gets right

For organizations that negotiate the right enterprise setup, OpenAI offers a recognizable ZDR path. That can work well if your engineering environment is centrally managed and developers use approved interfaces only.

In that scenario, you can standardize on:

• enterprise identity and admin controls
• approved local developer tooling
• explicit prohibition on cloud-agent execution for regulated workloads
• documented handling rules for PHI-adjacent code and test artifacts

That is a realistic operating model for mature teams.

The big risk with Codex

The largest risk is not the formal enterprise product. It is leakage into the wrong product tier or the wrong execution mode.

Two failure patterns stand out:

1. Developers use personal or consumer ChatGPT accounts for work tasks
2. Developers trigger cloud-hosted agent workflows assuming enterprise settings cover everything

In a HIPAA setting, both are serious problems. Consumer-facing ChatGPT tiers are not the same as an enterprise HIPAA-eligible workflow. And “our company has enterprise access” does not automatically mean every Codex path has zero-retention treatment.

Best fit for OpenAI Codex

Codex fits teams that already run a controlled enterprise environment and can sharply separate approved local tooling from unapproved cloud execution. It is especially suitable where the biggest operational challenge is user behavior, and the organization is prepared to enforce policy through endpoint management, access controls, and review gates.

It is a poor fit for bring-your-own-tool cultures. If developers frequently move between personal accounts, experimental agents, and sanctioned tools, the boundary is too easy to cross.

Why Vertex feels more predictable

Vertex AI’s appeal is structural. It gives teams a platform posture: cloud account controls, centralized logging strategy, policy management, and coverage that appears aligned with regulated workloads.

That can reduce ambiguity when compared with product-by-product negotiations. It is particularly useful for teams already standardized on Google Cloud and comfortable managing compliance through infrastructure rather than through isolated app-level promises.

There is also an important practical point: models accessed through Vertex follow Vertex’s governance context. If a team wants Claude-family capabilities but prefers cloud-level privacy architecture, that can be an attractive compromise.

The catch with Vertex

Vertex is not “zero retention everywhere by default” in a simplistic sense. The context provided notes that some model paths may use in-memory caching with a time-to-live for latency optimization. If that is enabled by default, regulated teams need to review it and disable it where required.

This is the typical Vertex trade-off. You get a more stable platform compliance posture, but you still need to understand configuration details. Architectural safety does not remove the need for implementation precision.

Best fit for Vertex AI

Vertex AI is the strongest fit for organizations that want privacy controls embedded in cloud architecture rather than negotiated tool-by-tool. It works well for platform teams that already manage regulated systems in GCP and want centralized control over routing, IAM, network boundaries, and model access.

It is less ideal for teams looking for a lightweight self-serve coding assistant rollout. Vertex is usually a platform decision before it is a developer convenience decision.

1. How you get protection

• Claude Code: through direct enterprise contracting, with scope limits
• OpenAI Codex: through enterprise approval, again with scope limits
• Vertex AI: through cloud platform commitments and BAA-backed infrastructure posture

If your organization wants a contract exception for a specific tool, Claude and OpenAI are the more direct comparison. If you want privacy controls rooted in cloud architecture, Vertex changes the frame.

2. What breaks the safe path

For Claude, the risk points include file retention, metrics logging, batch processing, code execution, and beta features.

For Codex, the risk points include cloud-hosted execution and any drift into consumer ChatGPT products or unsanctioned accounts.

For Vertex, the risk is usually configuration drift rather than consumer misuse. Caching, project settings, and governance discipline matter more than tier confusion.

3. Which model is easier to govern

Vertex is generally easier to govern at the platform level. Claude and Codex often require more precise “allowed path only” policy design.

That does not make Vertex automatically better. It means the control surface is different. Some teams prefer cloud architecture as the anchor. Others want direct provider relationships and are comfortable enforcing narrow usage patterns.

PHI-adjacent code is still sensitive

Many teams focus only on obvious patient records. But regulated exposure also shows up in:

• business logic reflecting care pathways
• eligibility windows or payer rules embedded in code
• stack traces containing patient identifiers
• fixtures and test datasets copied from production patterns
• support transcripts pasted into prompts during debugging

A tool does not need a full medical chart to create compliance risk. Small fragments can be enough.

Choose Claude Code if

• your team specifically wants an agentic coding experience
• you can negotiate and document the exact approved Anthropic path
• security can block non-approved APIs and features
• developers can be trained to avoid file-heavy and batch-heavy regulated workflows

This is the “high capability, high policy precision” option.

Choose OpenAI Codex if

• you already operate a strong enterprise workspace with tight admin controls
• local CLI and IDE workflows are enough for regulated tasks
• you can forbid cloud-agent paths for HIPAA-relevant work
• your bigger concern is shadow AI use, and you are ready to enforce sanctioned access

This is the “works if tightly managed” option.

Choose Vertex AI if

• your compliance posture is cloud-first
• your organization prefers structural privacy controls over product-specific ZDR language
• platform engineering already manages regulated workloads in Google Cloud
• you need consistency across model providers within one governed environment

This is the “architecture first” option.