What happened
OpenAI’s developer documentation introduces the CSAM Image Classifier as an image-screening tool for approved organizations. OpenAI says those organizations can use it at no cost. The stated purpose is not to replace a reviewer or make a final determination about an image; it is to help organizations prioritize images for human review.
The classifier returns two separate signals. One concerns potential child sexual abuse material, commonly abbreviated as CSAM. The other concerns potential AI-generated CSAM. OpenAI’s description therefore distinguishes between those potential categories rather than presenting one undifferentiated moderation signal. An organization using the classifier can treat the results as inputs to an existing review queue and decide how its human process responds to each kind of signal.
The boundary in OpenAI’s framing is important. A prioritization system changes which items a reviewer sees first; it does not, by itself, establish that a flagged image meets a legal or policy definition. The documentation describes human review as the destination of the workflow. This makes the product relevant to organizations that already have people and procedures for reviewing image content, especially where reviewers need help deciding what may require attention sooner.
Access is conditional on approval. OpenAI calls the eligible users “approved organizations,” so the offer is not described as an unrestricted public API release. The documentation’s concrete commercial claim is that approved organizations can use the classifier without charge. It does not turn that statement into a general claim that every organization, developer, or consumer account can immediately use the tool.
OpenAI’s published description identifies neither a model version nor a technical interface in the material at hand. It also does not state request or response formats, throughput, usage caps, an approval timetable, or a service-level commitment. Those omissions do not change the announced product boundary: free use for approved organizations, two separate signals, and a role in sorting work for people. They do mean that a team cannot infer integration effort, operating capacity, or expected results from the announcement alone.
For builders, this is best understood as a specialized moderation component rather than a general-purpose image classifier. Its value is potentially in queue design: using a signal to route potentially urgent material to trained human reviewers. The source does not prescribe the escalation policy, confidence threshold, retention practice, or reviewer safeguards that an organization should use. Those decisions remain part of the adopting organization’s operating model.
What the archive adds
New Runtime’s earlier coverage examined free access to OpenAI Auto-review checks for Codex users in an October 6, 2026 note. That change concerned a code-review approval workflow. This is a new event: a separate, image-focused classifier whose stated function is to prioritize human moderation work.
Trend signal
This does not yet fit a pattern New Runtime tracks. The announcement is a bounded moderation offer, not evidence by itself of a broader change in how agent systems expose capabilities, manage runtime security, or organize verification work.
Why it matters
Image moderation is partly a scheduling problem. Review capacity is finite, while some material can demand faster attention than routine content. OpenAI’s classifier is positioned to help an approved organization sort that incoming work by returning signals for potential CSAM and potential AI-generated CSAM before human review.
The free-access claim matters because it removes a stated usage price for organizations that meet OpenAI’s approval condition. But cost is only one part of deployment. A useful workflow still needs trained reviewers, a defined escalation path, appropriate access controls, and a method for assessing whether the tool’s signals improve prioritization rather than merely creating more work.
The announcement also reinforces a practical moderation principle: a model output is a routing input, not a final adjudication. Teams should not confuse a classifier’s flag with proof about an image. The work that follows the signal—review, documentation, escalation, and safeguards for staff—remains central to the system’s actual safety outcome.
Limits and uncertainty
OpenAI’s documentation does not establish the classifier’s accuracy, recall, false-positive rate, false-negative rate, or performance across image types and operating conditions. It does not quantify how the two signals should be calibrated, or identify an external evaluation of their reliability. The public description also leaves the approval process, eligibility criteria, usage limits, technical integration details, and service commitments unspecified.
As a result, the operational effect cannot yet be measured from OpenAI’s announcement. It is unconfirmed how quickly approved organizations can gain access, how the tool behaves at production scale, and which review procedures make its output useful. Organizations should treat the classifier as a prioritization signal and validate it within their own lawful, human-led moderation process.
Skill to build now
If your team already reviews sensitive image content, map the handoff before adding any classifier. Define which reviewers may see flagged items, which signals trigger faster triage, who owns escalation, and what record is kept of the human decision. Use separate handling paths only when the people, policy, and legal obligations differ; otherwise, preserve a single review queue with clear priority rules.
Do not set a confidence threshold or automatic enforcement action based on this announcement alone. First establish how a flag will be checked by a person and how false alarms will be recorded. That makes it possible to evaluate whether a screening signal reduces response time without weakening the review process.
Reader takeaway
OpenAI is offering approved organizations free use of an image classifier that produces separate signals for potential CSAM and potential AI-generated CSAM. Its stated role is to help people decide what to review first. The announcement changes access, not the need for trained human judgment, careful escalation, and local validation.
Open question
What evaluation results, approval criteria, and operational safeguards will OpenAI publish so organizations can calibrate these signals before relying on them in live review queues?