Tracking AI/SaaS pricing & policy changes: pre-launch field research

TL;DR

• Demand is real and measurable. Vendor pricing/limit changes repeatedly hit 300–800 HN points: GitHub Actions pricing changes 802pt/819 comments, Claude Code weekly rate limits 609pt/705, Cursor removing cost info from its usage page 337pt. HN stories matching "pricing" jumped from 682 per 12 months (Oct 2023–Oct 2024) to 1,238 (Oct 2024–Oct 2025, +81.5%), holding at ~1,217 in the latest 12 months.
• Not an empty market. AI API price history, SaaS price logs, deprecation trackers, ToS diffs — all four lanes already have occupants. 70+ similar services verified live during this research.
• What's left is the editorial layer. Most incumbents are auto-crawler snapshot sites or personal usage-tracker apps. Nobody found turns official-source diffs into edited, verified news — even after a deliberate falsification attempt (§3).
• The pipeline is not the moat. 15 official changelog RSS feeds and 13 statically-fetchable pricing pages confirmed. Collection is easy. Candidate moats: (1) a snapshot archive that gets more unique over time, (2) an accuracy brand, (3) the email list.
• Monetization: sponsorships first. Measured rates — DevOps'ish $750/wk (7k subs), Changelog News $1,500/wk (~27k). OpenAI/Anthropic have no affiliate programs; paid alerts are a $10–15/mo market — too cheap to lead with.

Method & verification

4 parallel research streams: competition sweep, demand measurement (HN API + search autocomplete), vendor-source live checks (57 URLs fetched), monetization benchmarks. Every claim requires a fetched URL; unverifiable items marked. Before merging, the parent re-verified headline numbers against original sources: 6/6 HN stories matched, 4/4 sampled competitors live with real content, RSS counts re-measured (3 matched; one was my own counting bug, corrected to the reported 138 items). Frequency numbers were re-measured with explicit windows rather than trusting the team's figures. One falsification probe actually changed a verdict — §3.

1. Demand

• "Pricing Changes for GitHub Actions" — 802pt / 819 comments (Dec 2025)
• AWS 15% GPU increase, quietly on a Saturday — 755pt / 481 (Jan 2026)
• "Claude Code weekly rate limits" — 609pt / 705 (Jul 2025)
• "GPT-5.6 Sol Pricing Cut by 50% on OpenRouter" — 636pt / 449 (Aug 2026)
• "OpenAI dropped the price of o3 by 80%" — 517pt / 495 (Jun 2025)
• TechCrunch: "Anthropic tightens usage limits for Claude Code without telling users" (Jul 2025) — the "quiet reduction" frame draws the biggest reactions
• "Cursor removed cost information from the usage page and CSV export" — 337pt / 153 (Aug 2026)

Recurrence: Claude Code limits alone produced 5+ large stories in 24 months. DDG autocomplete surfaces real search demand ("claude code usage limits reddit", "github copilot pricing change", "cursor rate limited by model provider").

2. Competition — all four lanes are taken

• A. AI/LLM API price tracking (most crowded): AI Change Watch (auto-detected provider changes with before/after + source links), AI Pricing Hub, LMMarketCap (350+ models), AI Pricing Guru, AICostSim, aichangelog.dev, Mungomash, 3 weekly price newsletters, and personal GitHub Pages daily crawlers. Snapshot comparators (OpenRouter, Artificial Analysis, LLM-Stats) and datasets (LiteLLM's 4,245-model price JSON) are upstream layers, not competitors.
• B. SaaS price logs (mature): PricingSaaS Pulse (claims 4,435 pages/week tracked, 204 changes published weekly), SaaS Price Hub/Pulse, PriceTrack, CompareTiers, SaaSBinder ("re-verifies vendor pages weekly"), SaaSTrueCost index.
• C. Deprecations: LLMStatus, AI Model Graveyard, quora/model-deprecation-tracker, DepWatch, DeprecationWatch, APIDrift. Big precedent: endoflife.date (Show HN 248pt).
• D. ToS diffs: Open Terms Archive, TOSTracker, Terms Watch, ShadyTerms. Paid precedent: TrackChangesHQ ($199 evidence packages).
• E. Adjacent — change-detection tools: changedetection.io (HN 253pt), Visualping, Distill — DIY substitutes.
• F. Commentary (different lane): Growth Unhinged (85k+), PricingBrew, Metronome Pricing Index.

3. The falsification probe

Trying to kill "subscription-tool limit tracking is an empty niche" surfaced OpenUsage (open-source macOS menu-bar usage tracker for Cursor/Claude Code/Codex/Grok), AI Limits Tracker, LimitTracker, and ccusage-style dashboards — the personal usage-tracker layer is already occupied. But all of these are personal dashboards. No editorial outlet covering the vendor's changes as news was found.

Updated verdict: "empty niche" is dead for tools (don't build a tool); alive for edited news/newsletters — with the usual absence-of-evidence caveat. Notably, the highest-reaction event types ("quiet reductions", transparency removals) are exactly what incumbents don't catch.

4. Sources are easy — so not a moat

57 vendor URLs fetched: official RSS on most changelogs (OpenAI 138 items, Anthropic 157 with next-day freshness, Cursor 50, Supabase 231, GitHub, Linear 253 + JSON, Notion 154, Vercel /atom, AWS What's New). No pricing page has a feed, but 13 of them parse fine with curl (OpenAI, Anthropic, Stripe, GitHub, Vercel, Supabase, Linear, Notion, Slack, Zapier, Cursor, Mistral, Datadog). Exceptions: ai.google.dev blocks bots behind Google login (proxy needed), HubSpot pricing is JS-rendered, AWS health exposes JSON instead of RSS (UTF-16). Watch the old URL: openai.com/api/pricing is 403; the canonical is developers.openai.com/api/docs/pricing.

Anyone can build this pipeline — which is why it isn't the moat. Candidates: an ever-growing snapshot archive (one well-known free-LLM repo 404'd during this research — data sources have short lives), an accuracy brand, the email list.

5. Monetization benchmarks (measured)

• Sponsors (1st): DevOps'ish $750/wk flat (7k subs, 46% open, 10% CTR — verified), Changelog News $1,500/wk no minimums (verified), Cooperpress public rate card (JS Weekly at 163k scale).
• Affiliates (2nd): Yes — Notion (up to $50 + 20% of Y1 revenue), Webflow (50% of first subscription × 12mo), HubSpot (30% recurring), monday.com (up to 100%), Jasper (25%), beehiiv (60%). No — Anthropic (404, verified), OpenAI (reported 404; re-check bot-blocked), Zapier. "Recommend AI tools for affiliate money" is half-true at best.
• Paid alerts (3rd): Visualping $14/mo, Distill $15–35, ChangeTower $12–78, Fluxguard $110 (25 sites) — too cheap to lead with; premium add-on only.

6. Pre-mortem

1. Losing to free (vendor RSS + changedetection.io) → sell summary + impact analysis, not detection.
2. SEO dependence → independent-site traffic collapse is documented (Housefresh, 84pt); own the email list first.
3. Affiliate single-point dependency → 37signals shut its program down.
4. Audience stuck below the sponsor threshold ("Ask HN: 5k subscribers, how do I get sponsors") → set kill criteria.
5. Bot-blocking breaks collection (403s observed live during this research) → diversify targets, prefer official feeds.
6. 2026 entry crowding (Competiflow, Adversa, Canary…) → be a publication, not a tool.

7. Verdict: conditional go

Reposition from "AI API price tables" to: subscription dev-tool (Cursor, Claude Code, Copilot, Codex) pricing/limits/policy/transparency changes, plus an integrated feed for SaaS price changes. Story format per demand evidence: (a) change summary, (b) before/after numbers with official diff links, (c) community reaction. Distribution: HN + newsletter first, SEO secondary. Revenue order: sponsors (5k+), then mid-SaaS affiliates, then premium alerts. Kill/pivot criteria: under 1,000 subscribers at 6 months or under 5,000 at 12 months.

The original idea's core survives — "not AI-written articles, but the fastest accurate digest of official changes" — with one correction: most incumbents are already auto-crawlers, so the edge is not speed. It's verification (linked diffs) and editorial judgment (impact analysis). And "the fact-check pipeline is the moat" should be revised: the pipeline is table stakes; the moat is the archive, the trust, and the list.

Caveats

Reddit measurement was blocked (403/429) — community demand rests on autocomplete evidence. Competitor traffic/revenue are private; self-claims quoted as-is. "Editorial gap" is an absence proof — X and the newsletter ecosystem weren't fully searchable. OpenAI's affiliate absence couldn't be re-verified (bot wall). A data source 404'd during this research; some links in this post will die too. That's the nature of this niche.