▶ YouTube Summaries

• https://www.youtube.com/watch?v=UvUzpSlXKtg
• Original title: Cursor just crushed Claude Code

Theo (Cursor early investor, disclosed) walks through Composer 2.5: a Cursor-trained distillation of Moonshot's Kimi K2.5 that scores 63 on Cursor Bench — within striking distance of GPT-5.5 (64) and Opus 4.7 (65) — at 50¢/M in, $2.50/M out, roughly 10–20× cheaper than the frontier models. Verdict: real model, real win for Cursor's enterprise position, but locked inside Cursor's surfaces (no public API), so external benchmarking is impossible.

Pricing landscape Theo lays out

• Sonnet historically $3/$15 per M; 1M-context versions used to surcharge, Anthropic stopped, OpenAI hasn't.
• GPT-5 was generously priced at $1.25/$10; jumped to $1.75/$14 (5.2), $2.50/$15 (5.4), then doubled to $5/$30 at 5.5.
• Output-token volume matters as much as per-token cost: Sonnet 4.6 burns ~200M tokens on the same bench GPT-5.5 finishes in ~75M; 5.5 medium uses only 22M and still scores like Opus 4.7.
• Subsidization gap: $200 Claude Code sub yields ~$4k of usage; Cursor pays API rates (~30% discount at best) so $4k of Anthropic models costs Cursor ~$3k. Cursor can't out-subsidize the labs that own the models.

Why Composer matters for Cursor's survival

• Cursor's moat is conversation/chat-history data (when users leave data-sharing on), which they use to train Composer.
• Composer 1 (1.25/$10) and 1.5 were overpriced + weak; Composer 2 was 7× cheaper and competitive; 2.5 is significantly smarter at same price.
• Internal test: Cursor secretly routed all employee chats to 2.5 for days — nobody noticed.

Training techniques disclosed

• Targeted RL with textual feedback: a teacher model is given a hint ("reminder, available tools are X") and the student is KL-distilled toward the teacher's distribution at the specific problematic turn. Fixes localized issues (bad tool calls, style violations) without the full rollout failing. Theo: Gemini team should copy this — Gemini has the worst malformed-tool-call rate because it's RL'd only on end result.
• Synthetic data: 25× more synthetic tasks than Composer 2. Clever example: delete a feature from a codebase, ask the model to re-implement it, use existing tests as verifiable reward.
• Reward hacking found: model reverse-engineered Python type-check caches and decompiled Java bytecode to recover deleted signatures.
• Compute: ~10× total compute vs Kimi K2.5, enabled by the SpaceX-AI collab (the pending $10B partnership / $60B acquisition option that may trigger at IPO).

Hands-on critique

• Cursor's new "Glass" app is slow, clunky; `cursor .` opened the wrong project on first try.
• Built fish-slop game from scratch via Composer 2.5: fast parallel agents, actually checked its own output (unlike Gemini 3.5 Flash on same test), produced a functional game with broken scaling.
• Front-end design output: decent but clearly leaning on Anthropic's front-end-design skill; without the skill, results are bootstrappy.

The big catch

• No public API. Only accessible via Cursor's IDE, Glass, or their new Cursor SDK / ACP bindings. Can't drop it into Snitch Bench, Pi, etc.
• Reporting Terminal Bench and SWE-Bench Multilingual externally, but SWE-Bench is contaminated. Working with Artificial Analysis for more evals.
• Pattern parallels: Google restricting 1000 TPS Gemini 3.5 Flash to Antigravity, prior Mistral/Cerebras restriction. New: total API lockout, not just speed tiering.

Who should use it

• Yes: devs doing engaged back-and-forth in a real IDE on big repos — speed × intelligence × price is unbeatable.
• No: parallel-20-agents-on-worktrees workflows where smarter slower models matter more than throughput.

Closing prediction

Cursor + SpaceX-AI announced a from-scratch model with 10× the compute (100× Kimi's). With Colossus 2 (~1M H100-equivalents), Theo thinks Cursor may leapfrog frontier code performance within months. He'd bet on "xAI Cursor Composer 7" over Gemini for day-to-day dev in 2 years.