Perspective API is shutting down: alternatives for toxicity and comment moderation

When does Perspective API stop working?#

Perspective’s own notice says: “The service will remain active until December 31, 2026.” It adds that usage requests and increased quota requests “will be addressed until February 2026”, and “We will not be offering direct migration support” [1]. The notice lists no attributes. Perspective is commonly described as returning 0-to-1 scores for attributes such as toxicity, severe toxicity, insult, profanity, identity attack and threat; Google’s model-card page did not render for us, so treat that attribute list as unverified.

Which alternatives exist, and how do they compare?#

Prices are as published on 10 October 2026, in USD, with the unit each vendor uses. Per-1,000 figures are not like-for-like because units differ (requests, records, characters). Where a vendor’s price was unclear to us, the table points to the vendor’s pricing page instead of a number.

Notes on caveats:

• OpenAI. The docs page states no terms about who may use the endpoint beyond one rule: do not send known or suspected child sexual abuse material [2]. Data goes to a US provider; check OpenAI’s current terms for your case. The endpoint does not target plain profanity as a category.
• Azure. The pricing page itself showed “$-” for the standard rate; the figure above comes from Microsoft’s public Retail Prices API and may differ by region [4][5].
• Google. moderateText also returns topic categories (Finance, Politics, Legal) that Perspective never scored, so a naive threshold will fire on unrelated content.
• Amazon. The pricing page’s rate for toxicity detection was not clearly readable on 10 October 2026, so we give no number; toxicity detection is not on the page’s list of free-tier APIs [8].
• Hive. Self-serve small customers are pointed to a separate vision-language-model product [10].

Where does StormDoor fit?#

StormDoor’s /v1/profanity and /v1/spam gates do two narrower jobs than Perspective did. They are not toxicity scorers: they do not rate harassment, hate, insult or threat on a scale, and nothing in them replaces Perspective’s TOXICITY or IDENTITY_ATTACK scores.

What they do cover:

• Word-list profanity with obfuscation handling. /v1/profanity matches whole words against per-language lists for 12 languages (en, sv, no, da, fi, de, nl, fr, es, it, pt, pl). It collapses repeated letters, folds common leetspeak and reconstructs spaced-out words such as f.u.c.k. Whole-word matching keeps place names like Scunthorpe clean. Text in a language without a list is not checked [17][18].
• Spam. /v1/spam runs rule checks (link density, invisible characters, mixed-script lookalikes, shouting, known spam phrases, off-platform contact invitations) and, on paid plans, a model classifier that emits spam.likely_spam as a separately labelled finding [19].

Price: the API is not yet available (waitlist). Planned launch pricing is 1 credit per profanity call and 3 credits when the spam model runs; Starter would be $19 per month for 25,000 credits. A one-time 5,000-credit trial is planned, with no free API plan. StormDoor has no customers yet and no published accuracy benchmark.

If your Perspective use was “block rude words in comments”, the profanity gate fits. If it was “score how hateful or harassing this is”, pick a scorer from the table.

How do I migrate off Perspective API?#

1. Map attributes. List which Perspective attributes your code reads. Map each to the closest category (for example INSULT to Comprehend INSULT or Google Insult; IDENTITY_ATTACK to OpenAI hate). Categories are not equivalent definitions.
2. Recalibrate thresholds. Do not reuse a cut-off such as 0.8. Score a labelled sample of your own comments with the new service and pick thresholds from the precision and recall you need. Azure returns severity levels, not probabilities.
3. Check language coverage. Count your traffic by language. Amazon’s documentation points to English only; Detoxify covers 7 languages; Azure names 8 tested languages.
4. Measure latency and limits. Hosted APIs add a network hop per comment; self-hosted models add GPU or CPU cost. Test p95 under your own load, and note rate limits (Azure F0 is 5 requests per second [3]).
5. Decide the failure mode. Choose what happens when the new service times out: publish unmoderated, queue for review, or reject.
6. Check data handling. Comments are user data. Confirm each vendor’s retention and sub-processor terms before sending them.
7. Run in shadow mode until 31 December 2026 so both systems score the same traffic.