Unmasking Cheetah – Inside Cursor’s Mysterious AI Model and Its Performance
Unmasking Cheetah – Inside Cursor’s Mysterious AI Model and Its Performance
Introduction
A new AI model called Cheetah has appeared exclusively on the Cursor platform, sparking curiosity among developers and AI enthusiasts. Marketed as a “stealth” model with a pay‑per‑token pricing structure, Cheetah’s capabilities and underlying architecture are not immediately clear. In this article we dive into the model’s pricing, test its performance on a series of agentic tasks, compare it with known contenders such as Claude, Sonnet 4.5, and GPT‑5 CodeX, and present the most plausible hypotheses about its true identity.
What Is Cheetah?
Cheetah is offered only through Cursor, an AI‑enhanced development environment. The model is described as a “stealth” offering, meaning it is not listed among the usual public models. Access requires a Cursor subscription, after which usage is billed at $125 plus $10 per million tokens for both input and output.
Key points about Cheetah:
- Exclusive to Cursor – not available via any other API or platform.
- Pay‑as‑you‑go pricing – a flat fee plus per‑token charges, similar to pricing for Gemini 2.5 Pro and GPT‑5 CodeX.
- Limited documentation – Cursor provides only minimal information, leaving the model’s internals opaque.
Pricing Context
The cost structure of Cheetah mirrors that of several high‑end models on the market:
- Gemini 2.5 Pro – identical per‑token rates.
- GPT‑5 CodeX – comparable pricing tiers.
- Claude Sonnet 4.5 – similar subscription levels for heavy users.
This pricing overlap suggests that Cheetah could be a repackaged version of an existing model rather than a completely new architecture.
Testing Methodology
To evaluate Cheetah, a series of agentic tests were run. These tests simulate realistic development tasks such as:
- Building a movie‑tracker app that uses Radix UI components.
- Creating a Kanban board with authentication and a database backend.
- Developing a Stack Overflow‑style Q&A site.
- Implementing an image‑cropping tool.
For each task, the model’s output was compared against results from Claude Code, Sonnet 4.5, and GPT‑5 CodeX. Success was measured by functional completeness, code correctness, and the ability to resolve errors autonomously.
Performance Comparison
Movie‑Tracker App (Radix UI)
- Claude Code produced a correct implementation with minimal errors.
- Sonnet 4.5 generated a similar response but displayed a known “Radix UI” error pattern.
- Cheetah produced a comparable solution but with more noticeable shortcomings, failing to resolve the Radix UI issue reliably.
Kanban Board with Auth & DB
- Claude Code delivered a fully functional prototype.
- Cheetah came close but left several bugs that it could not automatically fix.
- GPT‑5 CodeX also struggled, yielding incomplete code.
Stack Overflow‑Style Site
- Both Claude Code and Cheetah failed to produce a working version, indicating the task’s difficulty for current agentic models.
Image Cropper
- The task was a miss for Cheetah, Sonnet 4.5, and GPT‑5 CodeX, each returning non‑functional snippets.
Overall, Cheetah’s performance sits between Claude Code and Sonnet 4.5, often lagging behind Claude Code but occasionally matching Sonnet on simpler prompts.
Probing the Model’s Identity
A series of system‑prompt experiments were conducted to force Cheetah to reveal its internal name. The results were intriguing:
- When asked directly, Cheetah repeatedly identified itself as Claude.
- System instructions disclosed a generic description: “You are the mystery language model Cheetah by an unknown provider.”
These observations imply that Cheetah may be heavily fine‑tuned on Claude‑style outputs. Several hypotheses emerged:
1. A Grok‑Based Variant
- Grok Code exhibits similar behavior: it can be coaxed into breaking role‑play restrictions and then admits to being Claude‑like because of extensive Claude‑style training data.
- API logs from other services show Grok models appearing alongside GPT‑5 entries, supporting the idea that Cheetah could be a Grok‑derived model repackaged for Cursor.
2. Sonnet 4.5 Fast‑Edit
- Historically, Anthropic released a Sonnet 3.5 Fast‑Edit model to a limited group of Zed users. It was noted for high speed and near‑identical output quality to standard Sonnet.
- Cheetah’s speed and price resemble this fast‑edit lineage, but its functional gaps suggest it is not a true Sonnet 4.5 Fast‑Edit.
3. Custom Cursor‑Trained Model
- Cursor may have trained an in‑house model on Claude‑derived datasets, akin to the approach taken by other providers (e.g., WindSurf).
- However, achieving the nuanced similarity to Claude without direct licensing would be challenging, making this hypothesis less likely.
4. GPT‑5 CodeX Fast Variant
- Some API logs list GPT‑5 models near Cheetah’s entry, hinting at a possible fast‑mode version.
- Performance discrepancies, especially on complex tasks, argue against a straightforward GPT‑5 identity.
The most plausible conclusion is that Cheetah is a Grok‑derived model (potentially a successor to Grok Code Fast) that has been fine‑tuned on Claude‑style data, explaining both its pricing similarity and its tendency to claim it is Claude.
Implications for Developers
- Cost Transparency – The per‑token pricing aligns with premium models, so developers should budget accordingly for heavy usage.
- Performance Trade‑offs – While Cheetah offers rapid responses, its reliability on complex agentic tasks lags behind top‑tier models like Claude Code.
- Vendor Lock‑in – Since Cheetah is exclusive to Cursor, switching to an alternative platform would require re‑evaluating model choices.
Conclusion
Cheetah represents an intriguing addition to Cursor’s AI toolbox: a fast, stealth model priced like other premium offerings but delivering inconsistent results on demanding development tasks. Through systematic testing and probing, the evidence points toward Cheetah being a Grok‑based model heavily fine‑tuned on Claude outputs, rather than a genuine Sonnet 4.5 Fast‑Edit or a pure GPT‑5 variant.
For developers weighing AI model options, Cheetah may serve well for quick, low‑complexity code generation, but for mission‑critical or intricate agentic workflows, established models such as Claude Code or Gemini 2.5 Pro remain the safer bets.