monit.rs · Zenofex LLC

The September API drift report, corrected: we were counting the wrong thing

We found three defects in our own diff engine and re-ran September: 313 breaking changes became 1,905. But those are findings, not changes — the real figure is 248 distinct vendor actions, and it inverts the leaderboard.

The September API drift report, corrected: we were counting the wrong thing

Correction, 6 October 2026. This report originally said September produced 313 breaking changes. That number was wrong, and this page has been rewritten with the corrected figures. We found three defects in our own diff engine, fixed them, and re-ran every snapshot. The corrected count is 1,905 — six times higher. But the more useful discovery is that 1,905 is also misleading, and so was 313. Both count findings. The number of distinct things vendors actually did is 248. The rest of this report explains the difference, because it changes which vendors look worst.

TL;DR: 109 OpenAPI snapshots across 47 tracked public APIs in September 2026. 1,905 breaking findings, which collapse to 248 distinct vendor actions. Only 16 of the 47 vendors shipped a breaking change at all; 31 shipped none. By raw findings Vercel looks like the worst offender by a factor of six. By actions it is third, and the real answer is PostHog.

See it live: browse every detected change by vendor at monit.rs/changelog. No signup. Atom feed for all tracked APIs, or one per vendor.

Three defects, all in the same direction — the engine was blind rather than noisy.

Response severities were inverted. A field becoming optional in a response means the server may now omit something a client reads. We graded that safe. A field becoming required means the server now always sends it — strictly safer — and we graded that breaking. Both backwards.

Only part of each schema was examined. The differ walked properties and nothing else. Array items, allOf/oneOf/anyOf branches, additionalProperties and enum were invisible. Counting schema nodes across all tracked specs: 23,220 invisible against 15,329 visible. A list endpoint dropping a field from its element schema — GET /users returning [{id, email}] and losing email — produced no finding at all. That is the most ordinary breaking change an API can ship.

Request bodies were never diffed. We compared whether a body was required, never what was inside it. "This endpoint now requires a currency field" produced nothing.

We found the third one by running the differ against our own published OpenAPI spec and noticing it reported no change to something we had just changed. That is an uncomfortable way to find a bug in a product whose entire claim is detecting changes in other people's specs, and it is worth saying plainly rather than quietly shipping a fix.

The fixes added 238 findings in September alone of kinds the old engine could not express at all:

The part that matters more: findings are not changes

Re-running September took the count from 313 to 1,905. The temptation is to publish the big number. It would be dishonest.

A single edit to a vendor's spec can produce hundreds of findings, because most published specs inline the same schema at every site that uses it rather than referencing it once. One type, duplicated across four endpoints and two union branches, means one edit shows up eight times. Our stability scores have always collapsed this — scores are computed on distinct actions, not findings — but the reports were quoting the raw number. That was our error, not a change in the data.

The ranking inverts. By findings, Vercel produced 75% of September's breakage and the top three produced 91%. By actions, PostHog did the most distinct things, Cloudflare is second, Vercel is third with roughly half of PostHog's count, and the top three come to 67%.

Both columns are true. The left one measures how much of your integration surface moved. The right one measures how many decisions a vendor made. If you integrate with Vercel's web analytics, the left column is what hit you. If you are asking which vendor is managing their API carelessly, the right column is the fairer question.

Deepgram removed a single parameter. It appeared on 35 endpoints. That produced 49 findings with one distinct message between them.

Last month's report gave this its own section and called it a deliberate cleanup rather than a release that broke something. That read was right, and the deduplicated count now says so arithmetically: one action, not forty-nine. A vendor sweeping one deprecated parameter out of their spec should not outrank a vendor that made twenty separate decisions.

Nearly all of it is one event. On 27 September, Vercel's web-analytics response schema went from 236 properties to 21 — aiModel, asnName, appName and 212 others, gone. The same element schema is inlined across four endpoints, two of which carry two union branches, so 215 removed fields became 1,301 findings in a single snapshot.

We verified this before publishing, because a number that size is more often a bug than a fact: the removed field names appear nowhere in the new spec, there is no $ref indirection hiding them, and the schema shrank from 9,164 characters to 787. Vercel really did cut it.

This is also exactly the change the old engine would have missed entirely. It sits inside oneOf branches and array items — the surface that was invisible until last week.

The vendor whose number barely moves under deduplication. PostHog's 75 findings are 69 genuinely separate changes across their surface — almost no fan-out, almost no duplication. On the measure that asks "how many distinct things did this vendor change," PostHog was September's busiest API by a clear margin, and the original report ranked them first for the right reason even with the wrong engine.

Removals are 83% of all breakage. Things that were there and then were not.

Of the 1,905, 1,822 were on the response side and 83 on the request side. That split is itself new information — until this month we could not see the request side of a body at all, so the honest reading is that 83 is a floor rather than a measurement.

The practical point from last month's report survives unchanged, and the corrected data strengthens it: removal is the failure mode your monitoring is least likely to catch. A status check against a removed path returns 404, and plenty of uptime tools treat any response as "up." A schema check against a response missing a field returns 200 with valid JSON. The contract broke; nothing went down.

What this means if you consume any of these APIs

Check whether your vendors are in the 16 or the 31. Thirty-one of forty-seven shipped nothing breaking in September. "How often do APIs break?" is the wrong question; the distribution is nowhere near uniform.

Ask which number you are being shown. This applies to us and to anyone else publishing counts like these. A vendor with 1,431 findings and 32 actions is not six times worse than one with 231 findings and 66 actions — it is better by the second measure and worse by the first. Any report quoting a single number without saying which one it is has not done the work.

Assume removals. 83% of the risk is "this thing is gone," not "this thing changed shape." Test that the endpoints and fields you depend on still exist before testing what they contain.

Do not rely on the announcement. Every change here was found by diffing published specs, not by reading a changelog.

Every figure comes from diffing vendors' own published OpenAPI specifications on a schedule, snapshot against previous snapshot, with no editorial judgment in the counts. A change is breaking if it removes or tightens something an existing client could rely on.

"Distinct actions" collapses fan-out on two axes at once: one operation changing many fields, and one field changing across many operations. Per change kind we take the smaller of the two dimensions, so a parameter removed from 35 endpoints counts once, and an endpoint losing 215 fields also counts once. It is a deliberately conservative estimate of how many decisions a vendor made.

Three honest limits. We only see what a vendor publishes; an API with no OpenAPI spec is invisible to us, and a spec that lags its implementation gives a lagging signal. "Breaking" means breaking on the wire for some client, not for you specifically — whether a removed parameter matters depends on whether you were sending it. And we do not follow $ref into shared component schemas, which is the direct cause of the fan-out described above.

The corrected figures are live across the site: every changelog entry, every stability score and every badge was recomputed from the same re-diffed data, not edited by hand.

- All 47 tracked APIs - One Atom feed for everything, or per vendor at /changelog/<vendor.rss - Slack or Discord webhook on breaking changes

Monitor your own API this way

The figures on this page come from the same diff engine monit.rs runs against customer endpoints. Point it at your own API and it works the same way: it learns each response schema by example, diffs every subsequent response against that baseline, and alerts you when a breaking change ships. The free tier is 3 endpoints with the full feature set and no card required. Start monitoring free.

What monit.rs does

monit.rs compares observed response fields and types with a learned baseline. Run OpenAPI contract checks against the latest stored response when you want to compare the implementation with a declared contract.

See the difference when a field disappears or changes type. Optional AI summaries suggest possible causes and next steps; review the evidence before accepting a new baseline.

What you get on the free tier

  • Monitor 3 endpoints, full feature set, no credit card
  • Schema-aware regression detection — catches missing fields, type drift, new properties, status shifts
  • OpenAPI contract testing — validate the latest stored response on demand
  • Optional AI summaries — suggested causes and next steps alongside the observed diff
  • Public status pages at <slug>.status.monit.rs — share uptime history with customers
  • Multi-channel alerts — email, Slack, Discord, PagerDuty, webhook
  • Security fuzz testing — AI-generated adversarial edge-case tests to surface input validation gaps
  • MCP server for Claude Desktop / Cursor / Cline (no auth, three tools)

How it works

  1. Add your endpoint. Paste a URL. Set method + auth. Done.
  2. Baseline learned. monit.rs probes your endpoint and records the expected schema, types, and structure automatically.
  3. Regressions caught. Review detected changes and send a test alert to confirm your notification destination.

Public API changelog surface

monit.rs also runs a public changelog of 47+ well-known API vendors (Stripe, GitHub, OpenAI, Cloudflare, Datadog, PostHog, Vercel, Supabase, and 39 more). Every detected OpenAPI change is published at /changelog with per-vendor RSS feeds, stability badges, and side-by-side comparison pages.

See how we compare with existing monitors: vs Pingdom · vs Checkly · vs Uptime Robot · vs Statuspage.io · vs Bump.sh · vs Optic · vs Postman Monitors

Get started

Start monitoring free →

3 endpoints. Full features. No credit card required.