COMPARISON · PERSPECTIVE MIGRATION

Perspective is going away.
filter8() isn't.

Google's Perspective API — the free option a lot of teams leaned on — is being retired. Filter8 is the deterministic replacement: it returns a verdict you can act on, not a probability to threshold; it's sub-millisecond; and it keeps nothing after the response.

Try the playground → Read the docs

100 free calls · no credit card · one curl to your first verdict.

A verdict, not a score

Perspective hands you a probability and makes thresholding your problem. Filter8 returns allow / mask / reject / review plus clean_text — decided against per-rating policy.

Never guesses, never drifts

A hosted model can change under you. Filter8 is a pure function of (bundle, input) — same input, same verdict, forever, with a reproducible explain trace.

Your text, not theirs

No round-trip to Google. Filter every message at sub-millisecond speed, and keep nothing after the response — offset-only logs, no text.

Perspective vs Filter8, feature by feature

CAPABILITY	Google Perspective	Filter8
Status	Being retired	Actively developed
Approach	Hosted ML toxicity scoring	Deterministic engine — pure (bundle, input)
Output	Probabilities you must threshold	Verdict + clean_text
Latency	Network round-trip to Google	Sub-millisecond, in-process
Evasion (leet · homoglyph · spacing · zero-width)	Inconsistent on adversarial input	26 classes folded before matching
False-positive control (Scunthorpe)	Model-dependent	Explicit whitelist veto
Contact / PII classifiers	—	Phone · email · URL (incl. spelled-out)
Rating-aware (E–M · minors)	—	✓
Explainable / replayable	Score only	Full reproducible explain trace
Data residency	Text sent to Google	Nothing retained after the response
Throughput	Low default QPS	Your plan's rate
Price	Free — while it lasts	Free tier → flat, volume-priced

Comparison as of July 2026, based on public information about Perspective and Filter8's own positioning. Vendor capabilities change — verify specifics before relying on them.

THE SWAP

You were thresholding a score. Now you get an answer.

Perspective made you fetch a probability and decide what to do with it. Filter8 makes the call for you, against a rating-aware policy — one request, one verdict.

BEFORE · PERSPECTIVE

// a probability — you own the policy
const r = await client.comments.analyze({
  comment: { text },
  requestedAttributes: { TOXICITY: {} }
});
const score =
  r.attributeScores.TOXICITY.summaryScore.value;
if (score > 0.8) block();
// now what about masking? PII? multilingual?

AFTER · FILTER8

// a verdict + cleaned text, deterministic
const { action, clean_text, score } =
  await (await fetch(https://api.filter8.dev/v1/filter, {
  method: POST,
  headers: { Authorization: Bearer f8_live_… },
  body: JSON.stringify({ text, rating: T })
  })).json();
// action ∈ allow · mask · reject · review

WHERE PERSPECTIVE WAS STRONG

Honest about the trade

Perspective was genuinely good at nuanced, multi-attribute toxicity scoring across many languages — sarcasm, implied insults, coded language. A deterministic engine can't read intent, and we won't pretend otherwise.

That's exactly the review bucket. Filter8 handles the ~99% of traffic where speed, cost, privacy and precision matter deterministically, and flags the genuinely ambiguous tail so you can route it to a model asynchronously, off the hot path. Deterministic core, model escalation — the best of both, at a fraction of the model spend.

Perspective migration questions

Is Perspective really shutting down?

Per Google Jigsaw's own notices, it stops accepting new requests in early 2026 and goes out of service after. If you're on it, you need a replacement — that's the migration we're built for.

Is Filter8 a drop-in replacement?

The shapes differ in your favor: Perspective returns scores you threshold, Filter8 returns a verdict plus cleaned text. One call, and usually simpler code.

Perspective was free. What's the cost?

A free tier to start, then flat volume pricing well below per-call ML APIs. Our marginal cost is compute, so it doesn't tax every message.

Can it do multilingual toxicity scoring?

Filter8 is English-first and deterministic, so it doesn't emit nuanced multi-language probabilities. Route that ambiguous tail to a model async — deterministic on the hot path, model on the edge.

COMPARE MORE

vs bad-words — the npm wordlist → vs CleanSpeak — the enterprise suite → vs Azure Content Safety — hosted content safety →

Move off Perspective in an afternoon.

Start free, no card. Paste an evasion in the playground and watch it decide — same input, same verdict, every time.

Get an API key Open the playground