moderations`)

Status: DOCUMENTED + LIVE_VERIFIED (text and text+image with omni-moderation-latest). text-moderation-* RETIRED (removed 2025-10-27; live → 400). Sources: Moderations reference · Moderation guide · Pricing (Free) · Deprecations · Changelog (inline moderation scores) · model pages omni-moderation-latest, text-moderation-latest|stable · OpenAPI CreateModerationRequest / CreateModerationResponse. Last verified: 2026-09-18. Machine-readable: endpoints fragment, parameters/openai-moderations.json, objects/openai-media-objects.json.

# Models

Model	Inputs	Categories	Price	Status	Live
`omni-moderation-latest` (default) → `omni-moderation-2024-09-26`	text, `image_url` (URL or data URL, ≤ 20 MB); no audio	13 (incl. `illicit`, `illicit/violent`)	Free	DOCUMENTED, LIVE_VERIFIED	200 on text; 200 on text + 1×1 PNG data URL
`text-moderation-latest` / `-stable` / `-007`	text	11 (`illicit*` = null)	free	RETIRED 2025-10-27	`400 invalid_request_error param=model "Invalid value for 'model' = text-moderation-latest"`

Both omni-* ids appear in live /v1/models. Moderation results may also be requested inline with a generation: top-level moderation: {"model": "omni-moderation-latest"} in POST /v1/responses or POST /v1/chat/completions returns scores for the input and the output (after the full output when streaming; may hold an error object) — see the Responses/Chat agents' docs.

# Request

json

{"model": "omni-moderation-latest",
 "input": "a string"  |  ["several", "strings"]  |
          [{"type":"text","text":"…"}, {"type":"image_url","image_url":{"url":"https://… | data:image/png;base64,…"}}]}

One multimodal array = one combined result; an array of strings = one result per string (same order).

# Response and score semantics

json

{"id":"modr-6765","model":"omni-moderation-latest","results":[{
  "flagged": false,
  "categories": {"harassment": false, "...": false},
  "category_scores": {"harassment": 5.5e-05, "violence": 0.00054, "...": 0},
  "category_applied_input_types": {"harassment": ["text"], "violence": ["text","image"], "...": ["text"]}}]}

flagged: any category flagged (model-chosen thresholds).
categories: per-category booleans.
category_scores: confidence 0–1; not calibrated probabilities — OpenAI upgrades the model over time and warns that custom thresholds on category_scores may need recalibration.
category_applied_input_types: which input modalities contributed. Live: text-only input → every category ["text"]; text + image → image-capable categories ["text","image"], others ["text"]. Image-only input yields score 0 for text-only categories.

# Categories

Category	Description (docs)	Inputs
`harassment`	expresses, incites or promotes harassing language towards any target	text
`harassment/threatening`	harassment incl. violence or serious harm	text
`hate`	hate based on race, gender, ethnicity, religion, nationality, sexual orientation, disability, caste (non-protected groups → harassment)	text
`hate/threatening`	hateful content incl. violence/serious harm	text
`illicit`	advice/instructions for wrongdoing (e.g. "how to shoplift")	text
`illicit/violent`	illicit + violence or weapon procurement	text
`self-harm`	promotes, encourages or depicts self-harm (suicide, cutting, eating disorders)	text, image
`self-harm/intent`	speaker expresses intent/engagement in self-harm	text, image
`self-harm/instructions`	encourages or instructs self-harm	text, image
`sexual`	content meant to arouse; sexual services (excl. sex-ed/wellness)	text, image
`sexual/minors`	sexual content involving under-18s	text
`violence`	depicts death, violence, physical injury	text, image
`violence/graphic`	… in graphic detail	text, image

# Usage guidance (guide + safety best practices)

Free to use; recommended to screen both user input and model output; treat scores as signals for your own policy (filter, route to review, act on accounts), not as an automatic block; pair with safety_identifier (see docs/openai/safety.md). Images: URL or base64 data URL, ≤ 20 MB; audio not classified.

# Examples and tests

examples/openai/moderation/ — moderate.sh|py|ts (text) and moderate_image.py (multimodal data URL), all LIVE_VERIFIED. tests/openai/test_moderation.py (cheap, always runs).

# OpenAI Moderation API (POST /v1/moderations)