Visit NinjaChat: https://ninjachat.ai/
In this video, I’ll be telling you about Cursor’s new stealth model “Cheetah,” its pricing, real-world coding tests, and my best guess at which base model it actually is—plus how it compares to Claude Code, GPT-5 Codex, and others.
—
Key Takeaways:
🚀 Cheetah is a stealth model available only in Cursor, metered beyond the Pro trial/subscription.
💸 Pricing is $1.25/M input tokens and $10/M output tokens—similar to Gemini 2.5 Pro and GPT-5 Codex.
🧪 In agentic tests, Cheetah behaved very Sonnet-like (Radix UI error on Movie Tracker), hinting at its lineage.
🎮 It failed the Godot task (like Sonnet via Factory), while Claude Code uniquely succeeded there.
🧱 New tests: Svelte Kanban (auth + DB) nearly passed but got stuck; Nuxt Stack Overflow–style site failed; Tauri Image Cropper failed.
🤖 GPT-5 Codex also underperformed on these new benchmarks compared to Claude Code.
🔓 Prompting quirks let it drop the “Cheetah” roleplay and claim it’s Claude; system prompt calls it a “mystery model.”
🧩 Best guess: likely a Grok-based code model; alternatives considered include Sonnet 4.5 Fast Edit, a Cursor-trained model, or GPT-5 Codex Fast—unlikely to be Haiku.
👍 Overall: fast and interesting, but Cursor’s agentic stack trails Claude Code/Cline on complex tasks.
—
Timestamps:
00:00 – Introduction
00:07 – What is Cheetah and how to access it in Cursor
01:24 – NinjaChat (Sponsor)
02:14 – Test setup and evaluation criteria
04:54 – Prompting quirks: “Cheetah Stop” and identity admission & System prompt findings and implications
05:50 – Model guesses: Grok vs Sonnet Fast Edit vs Cursor vs GPT-5 Codex Fast
08:48 – Ending
source