We are currently building something exciting and will be launching soon. Registration is temporarily closed. Launching soon — registration temporarily closed.

Back to Blog

Perspective API Shuts Down After December 31, 2026: A Migration Guide

Google Jigsaw is retiring Perspective API: new access closed in February 2026, there is no migration support, and the service ends after December 31. A six-step checklist for the weeks that are left.

SE

Sentiqra Engineering

Product & Detection Systems

November 07, 2026
5 min read

Google Jigsaw is retiring Perspective API. New access closed in February 2026, there is no migration support, and the service ends after December 31. Here is what to do in the weeks that are left - whichever replacement you choose.


What Jigsaw announced

Perspective API has scored comment text for toxicity since 2017, and it sits inside a lot of comment sections, community tools, moderation queues and research pipelines. Its homepage now carries a sunset notice. As of November 7, 2026, it says:

  • The service stays active until December 31, 2026. After that the API is no longer in service, and access cannot be extended.
  • Usage and quota requests were accepted only until February 2026. If you are not already a Perspective user with enough quota, you cannot become one now.
  • There is no direct migration support. Jigsaw's reason: AI capabilities have evolved, and there is "less demand for a standalone tool specific to this area".

Source: perspectiveapi.com, checked November 7, 2026.

If Perspective is in your stack, the deadline is fixed. What is left to decide is the replacement and the order of the work.

Who is affected

  • Comment and community platforms that hide, hold or rank comments by TOXICITY score.
  • Moderation queues that use Perspective scores to decide what a human reviews first.
  • Plugins and integrations that call Perspective on your behalf. Check whether they have named a successor, or whether the feature simply stops working on January 1.
  • Researchers who measured toxicity or incivility with Perspective scores. After the shutdown, those scores can't be regenerated, so a study can't be re-run the same way.

Before December 31: a six-step checklist

1. Find every call site. Search code, configuration and vendor integrations for commentanalyzer.googleapis.com and the attribute names (TOXICITY, SEVERE_TOXICITY, IDENTITY_ATTACK, INSULT, PROFANITY, THREAT). Note which attributes each feature reads, and which thresholds it applies.

2. Archive what you will need later. For audits, appeals and research, store the scores you already have, together with the attribute, language, date and the threshold that was in force. Once the API is gone, "why was this hidden in March?" has to be answered from your own records.

3. Build a labeled evaluation set now. While Perspective still runs, sample real production text, including the hard cases: sarcasm, quotation, reclaimed slurs, and heated but legitimate argument. Label it with your policy team. This set is what you will judge every replacement against, and it outlives any vendor.

4. Write down what the replacement must do. At minimum, cover:

  • Languages. Perspective's TOXICITY supports 17 languages. Its other five production attributes support 7: German, English, Spanish, French, Italian, Portuguese and Russian.
  • Categories. Do you need more than toxicity? For example: threats versus insults, scams, harassment in direct messages, self-harm.
  • Output. One score per attribute, or ranked categories and labels you can route on.
  • Stability. Will the same text get the same result next month, or can the model change underneath you?
  • Data handling. Is submitted text retained? Where is it processed?
  • Cost, at your real volume.

5. Run the replacement in shadow mode. Send the same traffic to Perspective and the candidate, change no decisions, and log both. Compare the two on your labeled set, not on a vendor's benchmark.

6. Translate thresholds, then cut over with a fallback. A rule like TOXICITY > 0.7 → hold does not carry over one-to-one to a different output shape. Re-derive each threshold from your labeled set. Keep the old path as a rollback until the shutdown date.

Three weeks is enough for steps 5 and 6 if steps 1 to 4 are done first. Starting in November leaves a little margin. Starting in mid-December leaves none.

The lesson for whatever comes next

A shutdown like this is a reminder that your moderation results should not depend on one provider's continued goodwill. Whatever you choose, ask for things you can verify yourself:

  • A pinned configuration, so behavior does not change without your knowledge.
  • A way to check a past result without calling the vendor's live service.
  • A written retention and no-training commitment for the text you send.
  • An exit plan. Your labeled evaluation set and archived results are yours to keep, and so is any evidence attached to past decisions.

Where Sentiqra fits - and where it doesn't

Sentiqra is a text-analysis API. Its Trust & Safety analyzer returns structured categories and labels, each with a confidence level. The Perspective comparison page maps every Perspective attribute to Sentiqra categories. For example, INSULT maps to insults_derogation, and THREAT maps to threats_intimidation and incitement_violence.

What it adds:

  • Versioned, reproducible results. The same text under the same pinned configuration returns the same result. Each response carries a signed receipt that anyone can verify offline, which helps when a decision is appealed.
  • Zero retention. Submitted text is analyzed and not stored.
  • More than toxicity. There are separate analyzers for fraud and scams, dating safety, discrimination and more. For marketplaces and dating apps, see purchase- and romance-scam screening.

What to check before you switch:

  • Language coverage is different, not wider. Trust & Safety covers German, English, Spanish, French, Italian, Portuguese and Russian (all 7 languages of Perspective's other production attributes), plus Dutch and Polish, plus Finnish, Latvian, Lithuanian, Swedish, Turkish and Ukrainian, which Perspective does not offer. It does not cover Arabic, Chinese, Czech, Hindi, Indonesian, Japanese or Korean. If you rely on those, Sentiqra's Trust & Safety analyzer is not your replacement today.
  • The output shape is different. Perspective returns one probability per attribute. Sentiqra returns ranked categories and labels. Plan to re-derive thresholds rather than copy them.
  • Evaluate on your own data. Use the labeled set from step 3. The evaluation guide shows how to run the comparison locally, so your samples never leave your machine.

A shadow-mode sample of a few thousand messages fits in the free tier. Start with the quickstart.

Editorial note: This article was drafted with AI assistance and reviewed by Sentiqra Engineering for technical accuracy. Classification results shown here were generated using the pinned detection configuration identified above.