What Is Zero Data Retention?

Zero data retention (ZDR) is a commitment from a service provider, usually an AI model provider, not to store a customer’s prompts and outputs after a request has been processed. The data is used to generate the response and is not kept at rest afterward, except where the law or specific safety rules require it.

For AI code review, ZDR matters because review tools send source code to large language models (LLMs). A ZDR arrangement means the model provider doesn’t keep that code. It doesn’t, on its own, say what the review vendor stores, which is why teams need to ask about every layer.

What Changed in Zero Data Retention in 2026

Trend: model providers rework safety monitoring around ZDR

• OpenAI previewed Private Safety Processing to keep ZDR available for frontier models, running automated safety review on encrypted content held in customer-controlled storage, with no human access (August 19, 2026, OpenAI).
• OpenAI began rolling out Private Safety Processing to API customers in phases (September 2026, OpenAI).

Trend: ZDR scope narrows by feature and model

• Anthropic lists features that ZDR doesn’t cover and some models that need 30-day retention, so those models aren’t available under ZDR unless Anthropic authorizes it (last checked October 2026, Anthropic’s docs).

How Zero Data Retention Works

Under a typical ZDR arrangement with a model provider:

• Requests are processed and not stored. Prompts and responses aren’t kept at rest after the response returns.
• Abuse monitoring logs exclude customer content. OpenAI, for example, keeps abuse monitoring logs for up to 30 days by default and excludes customer content from them under ZDR (last checked October 2026, OpenAI’s docs).
• Stateful features are excluded. Features that store files, conversations or batches aren’t eligible, even when ZDR is on.
• Approval is needed. Providers enable ZDR per organization or project, after a request and review.
• Legal and safety exceptions remain. Providers keep some data where the law requires it.
• Gateways can require ZDR per request. Some model gateways route a request only to endpoints with a ZDR policy when the caller asks for it, and fail the request if none is available (last checked October 2026, OpenRouter’s docs).

Why Zero Data Retention Matters

Zero data retention limits how much sensitive data sits on someone else’s servers:

• Smaller breach exposure. Prompts and outputs that were never stored can’t leak in a provider breach.
• Less legal discovery risk. Data that doesn’t exist can’t be requested in litigation or by regulators.
• Easier privacy and compliance reviews. Data that isn’t stored is easier to square with rules such as GDPR, CCPA and HIPAA, so security, privacy and legal teams can approve AI tools more quickly.
• Less reliance on anonymization. Anonymized data can sometimes be re-identified, while data that is never stored can’t be.
• Protection for source code. Code is intellectual property, and it often contains secrets, internal URLs and customer logic.

Zero Data Retention vs Other Data Commitments

Zero Data Retention in AI Code Review Tools

An AI code review tool touches code in several places, and each one has its own retention terms:

• Model calls to an LLM provider, where ZDR applies.
• The review vendor’s service, which processes the diff and context.
• Codebase indexes and embeddings, which many tools keep so reviews can use full codebase context.
• Findings and analytics, which store snippets of code for dashboards and audits.
• Logs and telemetry, which may contain code or metadata.

A tool can have ZDR with its model providers and still store indexes and findings, often for good reasons. The question is whether the team knows and accepts what is kept.

Limits of Zero Data Retention

• ZDR covers one provider. It doesn’t extend to other vendors or tools in the chain.
• Some features fall outside ZDR. Stateful endpoints and some newer models may not be eligible.
• Data still leaves the organization. Code is sent to the provider for processing, even if it isn’t stored.
• Metadata is still kept. Providers record metadata such as timestamps, token counts and latency for billing and operations, and may keep classifier results about the content.
• ZDR breaks some features. Conversation history, memory and stored multi-turn state need data at rest. Under OpenAI’s ZDR, the store parameter is always treated as false, so stored responses aren’t available (last checked October 2026, OpenAI’s docs).
• Caching may still apply. Some providers keep repeated prompt content in a short-lived in-memory cache, which gateways may not count as retention.
• Tools and plugins have their own terms. Web search, file tools and other add-ons can be run by third parties with different retention policies.

Teams whose code can’t leave their network at all need self-hosted models and an on-premises review tool, as described on the on-prem code review page.

Questions to Ask an AI Code Review Vendor About Data Retention

1. Which model providers receive our code, and does each have a ZDR arrangement?
2. Does the vendor store code, diffs or embeddings, and for how long?
3. How long are logs kept, and do they contain code?
4. Is our code ever used to train models?
5. Can we bring our own model keys or self-hosted models?
6. Which deployment options exist: multi-tenant, single-tenant, on-premises or air-gapped?
7. Which certifications and reports, such as SOC 2 Type II, cover the service?

Qodo’s Take: Zero Data Retention at the Model Is One Layer, and Every Other Layer Needs an Answer Too

A security team approves an AI code review tool because its model provider has ZDR. Six months later, an audit asks where the codebase index lives and how long review logs are kept, and nobody has an answer.

Zero retention at the model is the goal. Zero retention everywhere would erase the audit trail, so governance needs some records, kept small and in a known place. Review tools should give a clear answer for every layer.

Qodo states its data commitments in public and lets teams choose where the whole review runs.

Qodo is the best AI code review platform for teams that need zero data retention because it answers for every layer: a zero data retention agreement with OpenAI for paying customers, no training on customer code, and on-premises installs that call the model endpoint the team chooses and store only the lines a finding refers to.

Qodo Data Commitments

Qodo’s enterprise page lists a zero data retention agreement, 48-hour log retention, no model training on customer code, SOC 2 Type II and encryption in transit (last checked October 2026, Qodo for enterprise).

The Qodo Trust Center describes this as a zero data retention agreement with OpenAI for paying customers (last checked October 2026, Qodo Trust Center).

Qodo’s pricing page lists strict data retention on its Pro Team and Enterprise plans and states that Qodo does not train models on customer code (Qodo pricing). Security documentation is available on request in the Trust Center.

Qodo Can Take the Model Provider out of the Chain

Qodo runs multi-tenant, single-tenant, on-premises or air-gapped, and Enterprise plans include bring-your-own LLM keys. On-premises, Qodo calls the model endpoint the customer chooses. With the team’s own gateway, no outbound access to api.openai.com is needed, though the default configuration calls OpenAI’s API (last checked October 2026, Qodo’s docs).

With self-hosted models such as NVIDIA Nemotron, code never reaches a third-party model API (October 1, 2026, Qodo 3.0 launch post). For the audit side, see SOC 2 and ISO compliance for code review.

Qodo Keeps Findings Records Small

Qodo keeps findings history for governance. In on-premises deployments with a findings database, Qodo stores only the lines a finding refers to, and no full file contents or complete diffs (last checked October 2026, Qodo’s docs).

Qodo Example: Answering the Security Questionnaire

This illustrative scenario is a composite, not a customer case study.

A healthcare software company sends its standard AI vendor questionnaire. For the multi-tenant option, Qodo’s answers cover its zero data retention agreement with OpenAI for paying customers, 48-hour log retention and no training. The company’s most sensitive repositories go to a Qodo on-premises deployment with a self-hosted model, so that code never leaves its network.

Ways to Limit AI Data Retention in Code Review

The table compares common ways teams limit what AI providers keep.

Qodo features in this table were checked against Qodo’s documentation and Qodo’s public website as of October 21, 2026.