PostHog Raised Its Entry AI Price 5.8x in One Day

Nine days ago I wrote that PostHog had clarified its per-event price at $0.00006 with the first 100k events free. On August 5 that line became $0.00035, and the single flat rate became a seven-band ladder.

Same method as always: every number below is a diff between two daily snapshots of a public page. Nothing here was announced.

How much did PostHog raise its AI observability price?

The entry rate went from $0.00006 to $0.00035 per event — 5.8x — for customers just past the free tier. The old rate still exists, but you now have to send 50 million events a month to reach it. The first 100k events stay free.

This is the ladder that replaced the flat rate, captured August 5:

• 100–150k events: $0.000350/event
• 150–250k: $0.000175
• 250–500k: $0.000100
• 500k–1 million: $0.000080
• 1–10 million: $0.000070
• 10–50 million: $0.000065
• 50 million+: $0.000060

The day before, there was one line under the free tier: 100k+ at $0.000060.

So what: if you budgeted AI observability on PostHog against the published rate, your number is wrong unless you send tens of millions of events. Price the band you land in, not the "from" figure at the top.

Why does the increase shrink as the customer gets bigger?

Because graduated bands stack the expensive rate at the bottom. Every account pays $0.00035 on its first 50k billable events, so a small account pays almost nothing but that rate while a large one averages it away. Run the monthly bill and the increase collapses as volume grows:

• 150k events/month: $3.00 → $17.50 (5.8x)
• 500k/month: $24.00 → $60.00 (2.5x)
• 1 million/month: $54.00 → $100.00 (1.9x)
• 10 million/month: $594 → $730 (1.2x)

Same page, same product, same day — a 5.8x increase for a side project sitting next to a 1.2x increase for an enterprise.

PostHog also launched Replay Vision in the same window, with 2,500 free credits (worth $25) in the free tier, and shipped 32 new MCP tools on August 3 including create, update and delete actions for Replay Vision. A new AI product line and a repriced AI product line, in the same week.

So what: "we raised prices" and "we raised prices on small accounts" are different decisions with different consequences. A graduated ladder lets a company do the second while the headline stays shaped like the first. When you model a competitor's pricing move, model the band, not the banner.

Did Cursor edit a changelog entry after publishing it?

Yes. On August 3 Cursor published Google Workspace Plugins, announcing that agents could read, write and act across Google Workspace, with named bullets for Google Drive, Gmail, Google Calendar, Google Docs and Google Sheets. By August 5 the Docs and Sheets bullets were gone and the summary had been rewritten to name only Gmail, Google Drive and Calendar.

The entry is still dated August 3. There is no correction note, no follow-up entry, and no mention of Docs or Sheets anywhere on the changelog today. Two integrations were announced as shipped and then removed from the record.

So what: a changelog looks like an append-only log and is not one. If you cite a competitor's changelog in a battlecard, a roadmap or a board deck, you are citing a mutable document. Screenshot it or diff it, because the version you read may not be the version that is there next month.

What changed on Stripe's pricing page four days running?

Stripe's pricing page produced a diff on August 3, 4, 5 and 6 — four consecutive capture days. Two of those changes matter; two are tests.

• August 4: stablecoins got an expiry date. The rate reads 0.8% per successful transaction, followed by "Promotional rate through January 1, 2027" and a further 0.2% thereafter. That is a 25% increase already on the calendar, published only on the pricing page.
• August 6: "Access to 100+" payment methods became "Access to 125+." A quarter more coverage in the headline count, no announcement.
• August 3 to 6: the per-product "Sign up with" call-to-action block appeared, was reshuffled, vanished, and came back. That is A/B testing on a pricing page, visible as churn.

Earlier in the same window Stripe introduced Radar Lite, an AI fraud-protection tier, and restructured Radar to offer both a monthly subscription and per-transaction pricing (July 29), then shipped API version 2026-07-29.dahlia with 30+ additions across payments, billing, invoicing and compliance (July 31).

So what: the stablecoin line is the one to write down. A promotional rate with a hard date is a price increase you already know about, and a 2027 model built on 0.8% is a quarter light. Nobody sends an email about the footnote under a rate.

Did Cursor change its price a seventh time?

No. After six changes in twelve days, Cursor's pricing page has not moved since July 28 — nine days. The last change there was structural rather than numeric: agent-usage multiples became the tier ladder.

It did tighten the free tier on July 29, limiting Grok 4.5 to medium effort and restricting Composer to non-fast mode, and it shipped iPad support across all paid plans on July 30.

So what: the $16 experiments stopped once the tiers were rebuilt around agent limits. That is the tell that the earlier flapping was a search for a number, and that the search has landed — at least for now.

What else moved across the stack

Windsurf spent the window on agent safety rather than features. v3.6.27 stops the edit, write, apply_patch and notebook_edit tools in Devin Local from writing through a symlink, so an approved edit can no longer be redirected to an unexpected file, and it restored loading Windows root and intermediate certificates so sign-in works behind a TLS-inspecting proxy or a corporate root CA. Before that, v3.6.21 and v3.6.22 added browser preview for remote agents, subagents in preview, editable command approvals and multi-user sessions.

Lovable shipped a per-user privacy control: any user on any plan can now exclude their own personal data from AI model training, with Business and Enterprise workspaces excluded by default and the toggle locked on. It also renamed its enterprise "private npm registry" to "managed registry" — same feature, settings label included — and added six connectors (Amazon Redshift, Microsoft Fabric, Google Analytics, Xero, Pipedrive, PostHog) plus a TanStack Start migration path.

beehiiv deleted one word from its pricing matrix: "Paid Recommendations" became "Recommendations", in the feature name and in all four plan rows. Availability did not change — Launch still no; Scale, Max and Enterprise still yes. When a company takes "Paid" out of a feature name across an entire comparison table, the commercial model behind the feature is usually what is being repositioned.

Kit rebuilt its resources hub around Industry Insights, retiring the Creator Stories, Customer Stories, Landing Pages & Forms and Reports groupings, and started leading with a "top-rated email marketing platform for creators in 2026" placement.

The pattern worth naming

Six companies, four categories, one window — and the thread running through all of it is agents getting write access to real systems, plus someone deciding who pays for it.

Cursor gave agents the ability to send mail and edit documents, then quietly pulled two of those integrations back out. Stripe prices a primitive for agents making purchases on a customer's behalf. Windsurf hardened the file-write path because an approved agent edit could be redirected somewhere it should not go. Lovable gave individuals a way out of model training. And PostHog repriced the layer that tells you what the agents actually did — most steeply for the smallest customers.

If you sell software, two things follow. The metered unit is moving toward agent activity, and the security review is moving with it. Both show up on pricing pages and changelogs weeks before they show up in a press release.

How this was made

Every fact above is a diff between two daily snapshots of a public page: pricing, changelog or blog. No press releases, no tips, no interviews. PostHog did not announce that its entry AI rate went up 5.8x. Cursor did not announce that Docs and Sheets came back out. Stripe did not announce that the stablecoin rate expires on January 1, 2027. Each of them changed a page, and the only way to know is to have kept yesterday's version.

One note on this edition, in the spirit of building in public: it is late. This teardown was due Monday. The snapshots kept running every morning and every number above was captured on schedule, but the AI layer that classifies them stopped working on August 1 and I did not notice until Thursday. Capture and analysis are separate systems here for exactly this reason — the data was still sitting there when I came looking for it. I have since added a step that re-classifies anything the AI missed, so an outage delays a briefing instead of quietly deleting a week of competitor changes.

If checking competitor pages by hand is not happening at your company — and it usually is not — here are the practical ways to track competitor pricing changes, including the automated one.

ScrapRival watches your competitors' pages and sends one briefing every Monday. You can read a real sample digest without signing up, or start a 14-day free trial pointed at your own competitive set.

Next Monday: whether beehiiv's dropped "Paid" turns into an actual pricing change, and whether Docs and Sheets come back.