Cursor Stopped Naming Composer on Its Paid Tiers
By Andrés Puentes González, founder of ScrapRival · How this is made
Last week I wrote that Lovable rewrote a published changelog entry to add AI model training on Free and Pro data, and promised two follow-ups: whether that September 9 date would survive the week, and whether beehiiv's Send API move onto Max would be followed by a price. Both are answered below.
Same method as always: 12 competitors, 24 public pages, one capture per page per day at 06:00 UTC, each compared to the previous morning's copy. Twenty-seven changes fired this week across 11 pages; 12 were scored relevant. Every quote below is a diff between two of those captures. Nothing here was announced.
And one section this week is a correction: three of the "deletions" my own tool reported were not deletions at all. That is at the bottom, with the number that explains it.
What changed on Cursor's pricing page?
On August 13, Cursor's pricing page lost a model name from three tiers at once. The feature line under Pro, Pro+ and Ultra read:
"✓ Generous limits for Grok & Composer"
By the next morning's capture it read:
"✓ Generous limits for Grok"
The same edit landed in all three paid individual tiers — Pro at $20/mo, Pro+ with 3x Pro limits, Ultra with 20x. Nothing else in those feature lists moved: "Access to frontier models", "MCPs, skills, and hooks", "Cloud agents" and "Bugbot on usage-based billing" are all still there, in the same order.
Composer has not been discontinued. As of today's capture the free Hobby tier still lists "✓ Access to Composer", and Composer still has its own entry in the site's Models nav and in the footer. What changed is narrower and more interesting: the paid tiers no longer name it as something you get generous limits of.
A second edit landed on the same page the same day. The Start plan — which the page says is available from July 28, 2026 — described what you get like this:
"Generous access to Cursor models: Grok 4.5 (fixed medium effort, non-fast) and Composer (non-fast), with enough usage to build with agents every day."
After the edit:
"Generous access to Cursor models: Grok and Composer, with enough usage to build with agents every day."
So what: both edits move in the same direction — away from specificity. "Grok 4.5, fixed medium effort, non-fast" is exactly the sentence you would quote in a procurement doc or a comparison table, because it names a version and two settings. "Grok" commits to nothing that a model swap could contradict. If you priced Cursor against a competitor in July using the model names on the page, the page you compared no longer exists, and nothing announced its replacement.
Did beehiiv rename Copilot to Agent?
Yes — and that is the whole change. On August 15, beehiiv's pricing page replaced every instance of one product name with another:
- beehiiv Copilot (Read Access) → beehiiv Agent (Read Access)
- beehiiv Copilot (Write Access) → beehiiv Agent (Write Access)
- The same swap in all four rows of the plan comparison matrix.
The "New" badge stayed. The gating is identical on both sides of the rename: Read Access is yes on Launch, Scale, Max and Enterprise; Write Access is no on Launch and yes on the other three. And the prices did not move — Scale is $43 per month, $517 billed annually, in the August 5 capture and in the August 15 one alike.
This also answers last week's open question. The Send API's move onto Max was not followed by a price change: ten days later beehiiv's whole price column is byte-identical. What changed instead was a word — the second rename on this page in two weeks, after "Paid Recommendations" became "Recommendations" on August 5.
So what: "copilot" was the 2024 word for a feature that suggests and waits. "Agent" is the 2026 word for one that acts. beehiiv shipped that repositioning without shipping a capability change or a price change, which is cheap and rational — but it means the word on a competitor's pricing page is now a weak signal of what the thing does. Judge the row, not the noun.
Why does Stripe's pricing page change every day?
Because it is being tested on, and this week removed any remaining doubt. Last Monday I called the flapping an A/B test and said the rule here is to rate-limit a page that alternates rather than report each flip. Stripe's pricing page then changed on all seven days:
| Date | What the capture showed |
|---|---|
| Aug 11 | Per-product "Sign up with…" buttons removed, replaced by "Start now" |
| Aug 12 | Guided discovery block back |
| Aug 13 | Discovery block removed, "Sign up with…" back, headline rate "2.9% + $0.30" added |
| Aug 14 | Discovery block back, "Sign up with…" gone, rate reads "2.9% + 30¢" |
| Aug 15 | Discovery block removed |
| Aug 16 | "Sign up with…" buttons back |
| Aug 17 | Buttons removed again, rate reads "2.9% + $0.30" |
Two things are being varied: the top-of-funnel treatment (guided discovery vs. direct per-product sign-up), and the rendering of the headline price itself — "$0.30" on the 13th, "30¢" on the 14th, "$0.30" again this morning. The underlying rate never changed. Its typography did.
So what: seven diffs, one fact — Stripe is optimizing the pricing page and has not touched pricing. A change detector pointed at this page would have sent seven alerts this week and been wrong seven times about their importance. This is the difference between detecting change and reporting it, and it is why the briefing is weekly.
Is Lovable's September 9 date still standing?
Yes. In the August 15 capture, Lovable's changelog still carries the entry rewritten on August 7, with the same sentence:
"Every user, on any plan, can now set their training-data choice in advance: starting September 9, 2026, Lovable may use customer data from Free and Pro plans for AI model training unless you opt out."
It still sits in the July 31 slot, still shows Jul 31, 2026, and still carries no correction note about the rewrite. Business and Enterprise workspace data remains excluded. Twenty-three days from the date it takes effect, nothing about it has moved.
The deletions that weren't
Three times this week my diffs showed entries vanishing from PostHog's changelog: an April 10 entry on August 13, a block of April 23 entries on August 14, a block of May 4 entries on August 15 — including "Together AI BYOK provider for LLM analytics" and "Resend source for data warehouse". After two weeks of writing about companies quietly deleting things from changelogs, that looks like a story.
It isn't one. PostHog's changelog page is 2,992,557 bytes of HTML. My fetcher caps a response at 1,500,000 characters. I have been capturing roughly the first half of that page, and every entry PostHog adds at the top pushes an older entry off the bottom of my copy — not off their page. The tell was visible before I went looking: on August 11 and 12 the last captured sentence ended at three different points mid-word, which is what a moving cut-off looks like and what a real edit never looks like.
So here is the rule I should have been applying, and now will be:
- A removal counts only when unchanged content sits below it. Cursor deleting Google Docs and Sheets from a published entry (last week) had untouched bullets underneath it. That is an edit.
- A removal at the very bottom of a capture, on a page that grew at the top, is a boundary artifact. That is my cap, not their decision.
The fetcher already returns a `truncated` flag when it hits the cap, and nothing downstream reads it. Raising the cap for long pages and refusing to report tail removals on truncated captures is the fix, and it is the next thing getting shipped here. Last week the blind spot was a classifier that under-weights deletions; this week it was a scraper that manufactures them. Both are worth writing down, because a competitive briefing whose failures you cannot see is not one you should be acting on.
What else moved
- PostHog published 21 changelog entries dated August 13 and 14 — 14 and 7 respectively: session-level AI Observability evaluations, Vercel AI SDK v7 support, nine new MCP tools, 13 more Reddit Ads warehouse tables, and audit-log export in OCSF format for SIEM ingestion. The Customer.io warehouse source left beta.
- Cursor shipped Builds — pre-staged cloud agent environments, quoted at 10x faster environment boot and 3x faster time to first token — and explicitly at no additional cost to existing Cloud Agents users.
- Lovable added Microsoft sign-in for end users' apps, dashboards built from Snowflake semantic views, and OpenAI image editing (GPT Image 2 and GPT Image 1 Mini) on top of generation.
- Loops shipped a Workflows API, a Claude connector and conditional sections.
- Windsurf shipped Devin v3.7.16 with conversation sharing on Devin Local.
The pattern worth naming
Not one of the 12 companies watched here changed a price this week. What moved was the vocabulary that describes what you are buying: a model name removed from three tiers, a version-and-settings clause replaced by a bare word, "Copilot" replaced by "Agent" across a matrix, and a rate rendered two different ways on alternating days.
Pricing pages are read as if they were specifications. Increasingly they are positioning surfaces that happen to contain numbers — and the words around the numbers change far more often than the numbers do, with no announcement attached to either.
How this was made
Every quote and number above is a diff between two daily captures of a public page, taken at 06:00 UTC and compared to the previous morning's copy, on the pages linked in each section. No press releases, no tips, no interviews. The correction section exists because the same method that catches a rewritten changelog entry also produces false positives, and publishing only the half that flatters the tool would make the other half unfalsifiable. The full method is on the methodology page.
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 Lovable's September 9 training default arrives on schedule, and whether Cursor's pricing page names a model again.