Methodology & data sources

Every tracked change links a primary source. We prioritise, in order: official docs, changelogs, pricing pages, and status pages maintained by the tool maker. Where official records are absent or retrospective, we use Wayback Machine snapshots to establish dates. Reputable technology news coverage fills the gap for announcements that lack a canonical permalink.

Community sources — forums, X/Twitter threads, Reddit — are used only as secondary confirmation, never as sole evidence. Events sourced this way are marked secondary source or unconfirmed inline.

289 tracked changes across 30 AI tools. 72% of those changes are sourced from official materials.

No claim appears without a source. Every event’s date, direction and figures come from its linked primary source. The longer written analysis is drafted with AI assistance from that sourced record and reviewed before publishing — it interprets the data, it never invents it.

This site is rebuilt automatically, and its data is reviewed against live sources on a rolling basis rather than on a fixed daily promise. Each tool record carries a last checked date that reflects when the current-state data was last verified against live sources — not just when the page was last generated. That date is shown on every tool page in the header and in the provenance strip beneath the title, so you can always see how fresh a given tool’s record actually is.

Timestamps are shown on every page and in the RSS feed. The most recent verification across all tools was 2026-07-23.

Every event carries one of three confidence levels. Events with no marker are high-confidence — that is the clean, default state.

High

Sourced from official documentation or a dated official announcement. Date, scope, and direction are confirmed. No marker shown — this is the default expectation.

Secondary source (marked inline)

Sourced from reputable news coverage, the Wayback Machine, or a status page without an accompanying official changelog entry. The event occurred but the primary record is indirect.

Unconfirmed (marked inline in italic)

Sourced from community reports — forum posts, social threads — or where the date or scope is approximate. Included for completeness; treat with appropriate scepticism.

Drift score — a number from 0 to 100, where 100 means no tightening pressure in the window. It starts at 100 and each tightening event subtracts a recency-weighted penalty. Older events decay toward a floor over roughly 12 months, so a change from a year ago matters much less than one last week.

Key multipliers applied before weighting:

• Tightening events count approximately 4× what loosening events add back.
• Silent (unannounced) changes carry an additional ~1.5× penalty.
• Loosening adds back, but asymmetrically — a full reversal does not fully restore the score.

Verdict — Stable / Watch / Caution — is derived from the recent trajectory (90-day delta and trend direction), not the absolute score alone. A tool can score well historically but earn a Watch rating if things have tightened lately.

Sub-scores (Pricing, Limits, Models, Transparency) apply the same recency-weighted logic scoped to the relevant event dimension. They are independent slices, not a weighted average of each other.

These are transparent heuristics over sourced events, not opinions — and explicitly NOT capability or quality benchmarks.

The User signals section on each tool page surfaces what people are actually complaining about. It is drawn from three free, public sources over the trailing 180 days: Hacker News (via its keyless Algolia search API — stories and comments mentioning the tool), GitHub issues (for tools with a public issue tracker, via the REST search API), and Apple App Store reviews (for tools with a consumer iOS app, via Apple’s public reviews RSS feed). Each match is run through the same keyword classifier and bucketed into a complaint category — limits/quota, pricing, model removal, performance, or other.

We show the live counts only when there are at least 5 matching complaint mentions across all sources; below that threshold we fall back to hand-picked, sourced quotes rather than dress up a tiny sample. Every mention is tagged with its source (HN / GH / App Store) and links back to the original thread, issue, or review so you can read the context. Not every tool has every source — only tools with a public GitHub tracker or a consumer iOS app contribute those signals. Reddit and X are deferred for now (no stable free API).

These are raw counts of public posts, issues, and reviews matching complaint keywords — not a scientific survey or a representative sample of all users. Percentages are shown only when the sample is large enough (≥10) to be meaningful. No count is ever fabricated; every number traces to fetched Hacker News, GitHub, or App Store data.

This tracker covers one thing: how AI tool constraints have changed over time.

• ✗Not model quality. Drift score says nothing about output quality, reasoning ability, or benchmark performance.
• ✗Not uptime or reliability. We do not track incident frequency or SLA adherence.
• ✗Not a ‘best tool’ ranking. A high Drift score means consistent, stable constraints — not that the tool is superior for your use case.

We track only: usage limits, quota changes, pricing moves, model access changes, and constraint policy shifts — and how each has evolved over time.