How Cursor launched Composer 2 on X

What did Cursor post?

Composer 2 is now available in Cursor.

That's the entire post: seven words, one image, no link in the text.

How was the launch post built?

The hook is flat by design: a statement of fact, not a pitch. No adjective, no benchmark, no "we're excited to announce." Composer 2 is named once and the sentence ends.

Format. A single image (2,400 by 1,800 pixels), 38 characters of text, no thread, no video, no carousel, no link.

Timing. The post went up Thursday, March 19, 2026, at 16:30 UTC, the same day Cursor's changelog entry for Composer 2 went live, describing it as delivering "frontier-level coding performance with strong results on challenging coding tasks." That page, not the post, is where pricing lives: $0.50 per million input tokens and $2.50 per million output tokens on the standard tier, $1.50 and $7.50 on the faster default tier. A model launch this technical could have become a ten-post thread; instead Cursor pushed every follow-up toward a page built for exactly that.

How did the launch post perform?

The post has 5.4M views, 9.6K likes, 863 reposts, 638 replies and 1.2K quotes, with 2K bookmarks.

Engagement runs at 0.23% of views, #43 of 76 by engagement among the launches tracked in Goated's launch directory, while by views it ranks #40 of 76. The post won on reach rather than reaction, built on one line and a single account's own audience, no video, no thread, no cast of accounts quoting it.

What happened around the post?

Composer 2 is Cursor's own proprietary coding model, not a wrapper around GPT or Claude, per SiliconANGLE's coverage. On CursorBench, it scored just over 60%, placing third behind two configurations of GPT-5.4, while beating GPT-5.4's lower-tier configuration and Claude Opus 4.6. On Terminal-Bench 2.0, which tests AI agents on realistic, multi-step command-line tasks, it also beat Anthropic's model, per Snorkel AI. The model supports prompts up to 200,000 tokens.

None of those figures appeared in the X post; they lived on the changelog and in outside coverage. SiliconANGLE also noted Cursor trained Composer 2 on coding-only data and built in a self-summarization step that compresses earlier parts of a session so the model keeps working inside its token limit on long tasks, a detail that would have sunk a one-line post but fits next to the pricing table on the changelog.

What can another team copy?

1. 1

   Write the post as a statement, not a pitch

   "X is now available" beats "We're thrilled to announce X." Save the enthusiasm for people who already want the news.

2. 2

   Publish the deep version somewhere durable, the same day

   Cursor's changelog carried the benchmarks and pricing before any outlet wrote about them, giving press one canonical page to link to.

3. 3

   Attach one piece of media, not a set

   A single image matched the weight of a one-line claim. A carousel or demo video would have asked for more attention than the post was spending.

4. 4

   Keep specifics off the post entirely

   Pricing, benchmark scores and context-window sizes belong on a page built to be read slowly, not inside a 280-character post.

See the live numbers behind this one on the Composer 2 launch page, compare it to every other launch Goated tracks in the launch report, or read the full playbook for launching a product on X.