What was text-moderation-stable?
text-moderation-stable was an OpenAI model alias for detecting potentially harmful content in text. It was not a general-purpose language model and was not intended to write answers, summarize documents, generate code, or take actions. Its job was narrower: evaluate text and provide moderation signals that an application could use in its own safety workflow.
The stable alias resolved to the text-moderation-007 model snapshot. OpenAI offered moderation models through the Moderation API at no charge. The model was useful for screening user submissions, checking generated text, and routing questionable content for additional review.
How the model classified text
An application submitted a text string, or an array of text inputs, to OpenAI's /v1/moderations endpoint. The response included an overall flagged decision, category-level flags, and confidence scores. Categories covered areas such as harassment, hate, violence, sexual content, and self-harm.
A flag was a signal for application policy, not a complete decision about what a user should be allowed to do. For example, a service might use a high-confidence result to block a post, send an uncertain result to a human moderator, or apply a separate rule based on the user's history and the surrounding conversation. The model's scores therefore worked best as one component in a broader review process.
Capabilities and supported modalities
text-moderation-stable accepted text input and returned structured text-based moderation results. It did not accept images, audio, or video. It also did not produce images, audio, video, embeddings, executable actions, or tool calls. Its output was a classification-style response containing moderation categories, flags, and scores.
This text-only design was an important limitation. A service receiving image captions, uploaded photographs, voice recordings, or videos could not send those raw media types to this model for direct assessment. Developers needed a separate workflow for non-text content, or a newer moderation model with broader modality support.
Technical specifications
| Provider | OpenAI |
|---|---|
| Model alias | text-moderation-stable |
| Underlying snapshot | text-moderation-007 |
| Primary task | Text safety classification |
| Input | Text |
| Output | Structured moderation categories, flags, and confidence scores |
| Knowledge cutoff | September 1, 2021 |
| Maximum output tokens | 32,768 |
| Context-window value | Not separately provided in the supplied documentation |
| API price | Free through the Moderation API |
| Deprecation announcement | April 28, 2025 |
| Access ended | October 27, 2025 |
The documented maximum output-token figure should not be interpreted as a promise that ordinary moderation responses would use that many tokens. Moderation responses are structured classification results, and the supplied model information does not provide a separate context-window specification.
Strengths and trade-offs
The model's main strength was specialization. Because it focused on text safety classification rather than open-ended generation, it fit applications that needed a straightforward moderation check. Its free API pricing also made it practical for legacy systems that needed to screen substantial volumes of text without adding a separate per-token moderation charge.
It was also easy to reason about operationally: submit text, inspect the category results, and apply application-specific rules. The returned category scores could support different thresholds for different risks instead of forcing every category into the same yes-or-no policy.
Its main trade-off was limited scope. It was text-only, tied to an older model snapshot, and designed for a narrower moderation task than newer multimodal systems. It was not a replacement for human review, identity or abuse controls, rate limiting, account enforcement, or policy logic. OpenAI also stated that the Moderation API was not designed to handle known or suspected child sexual abuse material and should not be treated as a dedicated child-safety system.
Reasoning, coding, and tool support
text-moderation-stable was not a reasoning or coding model in the usual sense. It did not provide general chain-of-thought reasoning, code generation, code execution, browsing, function calling, or external tool use. Any apparent sophistication in its category scores should be understood as classification behavior rather than a general reasoning capability.
For an application, this distinction matters. A moderation result could help decide whether another model's generated answer should be shown, revised, or reviewed, but text-moderation-stable could not itself rewrite the answer, explain a policy to the user, or investigate additional context through a tool.
Pricing and availability
When it was available, text-moderation-stable was free through OpenAI's Moderation API. That pricing applied to the moderation endpoint rather than representing a general free tier for OpenAI's other models or services.
The alias was deprecated on April 28, 2025, and access ended on October 27, 2025. As a result, it is a retired model rather than a viable choice for a new production integration. Existing systems that still refer to the alias should be migrated and tested against the recommended replacement instead of assuming that the old behavior remains available.
Replacement: omni-moderation
OpenAI recommended omni-moderation-latest as the replacement. The replacement belongs to a newer multimodal moderation direction and supports text and image inputs, whereas text-moderation-stable handled text only. The two models should not be treated as behaviorally identical: a migration can change category decisions, thresholds, response details, and handling of edge cases.
OpenAI also has a dedicated page for Omni Moderation, which is useful when evaluating the successor's broader input support. Applications should compare the replacement on representative content, calibrate policy thresholds, and confirm how the new response format fits existing enforcement logic.
When to choose text-moderation-stable
For a current project, there is no supported reason to choose this model because access ended in October 2025. Its historical role is still relevant when maintaining or auditing a legacy integration. In that context, the model's profile was a good fit for:
- Legacy text-only applications that needed a basic harmful-content classification step.
- Systems that relied on category flags and confidence scores for automated routing.
- Historical analysis of decisions made by the
text-moderation-007snapshot. - Documentation or migration work involving older OpenAI Moderation API integrations.
It was not an appropriate choice for new applications, direct image moderation, audio or video moderation, general-purpose reasoning, code generation, or workflows requiring current model support. New systems should evaluate omni-moderation-latest or another currently supported moderation option, particularly when content can include images as well as text.
Bottom line
text-moderation-stable was a focused, free, text-only OpenAI moderation model that returned category flags and confidence scores through the Moderation API. Its specialization and cost made it useful for earlier text-screening workflows, but its older snapshot, lack of multimodal input, and absence of general-purpose generation or tool capabilities limited its scope. Because OpenAI retired it on October 27, 2025, its practical importance today is mainly as a legacy model that must be replaced rather than deployed.