Cursor Composer 1 & SWE-1.5 (Fully Tested): LOL, HOW CAN A $10B COMPANY MAKE A MODEL THIS BAD?

Key Concepts

• Cursor Composer: A new model launched by Cursor, advertised as a frontier model for agentic coding, four times faster than similarly intelligent models, with low latency and aiming for sub-30-second completion times for most turns.
• SWE 1.5: A new model launched by Windsurf (or Cognition), claimed to be faster than Cursor's model, capable of 950 tokens per second, powered by Cerebras, and trained on an undisclosed open model with proprietary RL data.
• Agentic Coding: A type of AI-assisted coding where the AI acts as an agent, capable of performing multi-step coding tasks and iterating quickly.
• Open Weights Model: A model whose weights are publicly available, allowing for community inspection and fine-tuning. The transcript suggests both Cursor Composer and SWE 1.5 are based on Mistral 4.6.
• RL Data (Reinforcement Learning Data): Data used to train models through trial and error, rewarding desired behaviors and penalizing undesired ones.
• Leaderboard: A ranking system used to compare the performance of different AI models on various coding tasks.
• White Labeling: The practice of taking a product or service developed by one company and rebranding it as one's own.
• Photogenius AI: The sponsor of the video, an AI-powered creation suite offering image, video, and 3D model generation tools, including access to models like Google's Nano Banana and V3.

Model Performance and Criticisms

Cursor Composer

• Claimed Capabilities: Frontier model, four times faster than similarly intelligent models, designed for low latency agentic coding, completing most turns in under 30 seconds. Early testers found it delightful for quick iteration and trusted it for multi-step tasks.
• Observed Performance:
  • Movie Tracker App: Not great, with errors on some pages and a poor discover view. The reviewer states cheaper models can produce better code.
  • Go UI Calculator: Works pretty well and without issues.
  • Godo Game: Does not work, with Minimax and GLM performing significantly better.
  • Open Code Big Task: Does not work.
  • Spelt App: Only provides a login screen; the backend fails with numerous errors.
  • Next App: Does not even run.
  • Tari Rust Image Cropper: A significant failure.
• Leaderboard Position: Scored 11th position.
• Comparison to Other Models: Performance is very similar to Quen 3 coder and GLM. Kilo is significantly better, and even Minimax performs much better.
• Potential Base Model: The reviewer speculates it might be a GLM4.5 (not the latest 4.6 variant) or a Quen 3 coder finetune, leaning towards Quen 3 coder due to its slightly better performance than Quen 3 coder but not as good as GLM 4.5.
• Criticism: The reviewer expresses strong disappointment, calling it a "trash model" and questioning Cursor's ability to select a good base model for training, even with their significant company valuation. They believe RL won't help without correct pre-training alignment. The speed is deemed irrelevant if the output is poor.

SWE 1.5 (Windsurf/Cognition)

• Claimed Capabilities: Faster than Cursor's model, capable of 950 tokens per second, powered by Cerebras. Trained on an undisclosed open model with proprietary RL data.
• Observed Performance:
  • General Performance: Described as "even worse" than Cursor Composer.
  • Leaderboard Position: Scored 19th position.
  • Task Completion: Cannot complete tasks effectively.
  • Calculator: Makes a calculator but cannot perform calculations, deemed "kind of bad."
• Criticism: The reviewer finds it "pretty bad" and questions the strategy of these model companies, suggesting they should either use open models directly or hire proper ML teams to build from scratch. They are accused of being lazy, wrapping existing models, and forking VS Code. The model's performance is described as "trash code."

Sponsorship: Photogenius AI

• Description: An all-in-one AI-powered creation suite.
• Key Features:
  • Image Generation:
    • Google's Nano Banana for fast, high-quality image generation.
    • Ability to add reference images and perform edits within the tool.
    • Supports Flux, Stable Diffusion, Kandinsky, and more.
  • Video Generation:
    • Google V3 support with and without reference images.
    • Rendering in different styles without complexity.
  • 3D Model Generation:
    • Upload a PNG (e.g., Lego build, simple robot) to get a printable model.
    • Described as cheap, quick, and surprisingly clean for rapid prototyping.
  • Pricing: Among the best for V3 and Nano Banana.
  • Additional Tools: Access to about 10 other AI tools like avatars, background removal, logo, emoji, ads, and app icons.
• Pricing and Discount: Starts at a low entry price. An additional 30% off is available with the coupon code "king30."
• Call to Action: Check Photogenius AI via the link in the description.

Methodologies and Frameworks

• Testing Methodology: The reviewer tested Cursor Composer using the Cursor CLI tool and SWE 1.5 using its editor interface. Both models were found easily within their respective interfaces.
• Performance Evaluation: Models were evaluated based on their ability to complete various coding tasks, including building specific applications (movie tracker, calculator, game), handling backend logic, and image cropping. Leaderboard positions were used as a quantitative measure of performance.
• Underlying Model Speculation: The reviewer attempts to infer the base models used by Cursor Composer and SWE 1.5 based on their observed performance and known capabilities of other models like GLM4.5, Quen 3 coder, and Mistral 4.6.

Key Arguments and Perspectives

• Lack of Credit to Open Models: A central argument is the ethical concern and lack of transparency regarding both Cursor and Windsurf not crediting the open-weight models (likely Mistral 4.6) they are based on. The reviewer finds this "insane" and suggests it's akin to "ripping off an open model for profit."
• Disappointment with Commercial Models: The reviewer expresses significant disappointment with the performance of both Cursor Composer and SWE 1.5, especially after testing superior open-source models like Minimax. They argue that these companies are not innovating but rather "wrapping something and forking VS Code" or "white labeling."
• Focus on Speed vs. Quality: The reviewer criticizes the emphasis on speed without delivering quality code. They argue that faster generation of "trash code" is not valuable and that speed can be achieved with other, better-performing models or even diffusion LLMs for tool calls.
• Call for Transparency and Proper Development: The reviewer advocates for either directly using and open-sourcing good open models or for large companies to invest in proper ML teams to develop models from scratch.
• Critique of Cursor's Business Model: The reviewer questions how a company valued at $10 billion can fail to produce a good model, even with abundant open-source options.

Notable Quotes and Significant Statements

• "It's insane how both of these companies don't give any kind of credit to the model they've trained on and people are somehow fine with it." (Reviewer)
• "I would recommend not using these models just based on that fact alone." (Reviewer, referring to the lack of credit)
• "The cursor model's movie tracker app is not great. It has errors on some pages. The discover view doesn't look very good. And overall it's not a great experience. Cheaper models than this can make better code." (Reviewer)
• "Both of them are just bad, and I wouldn't recommend anyone to use them at all." (Reviewer, concluding on Cursor Composer and SWE 1.5)
• "It's also insane to me how people are not criticizing both these companies for not giving credit to the open model used underneath." (Reviewer)
• "If they're basically ripping off an open model for profit and not even open sourcing it, then they should at least credit the model and disclose which one they used." (Reviewer)
• "I mean, it's fast, but it generates trash code. So, I think it's fast at generating trash code if that's what you care about." (Reviewer, describing SWE 1.5)
• "Deploying an open model with a bit of RL and fine-tuning, not giving credit, and deploying it on Cerebras or Gro while calling it your own. That's not a good move." (Reviewer)

Technical Terms and Concepts Explained

• Mistral 4.6: A specific version of the Mistral AI large language model, likely serving as the base for Cursor Composer and SWE 1.5.
• Cerebras: A company that provides specialized hardware (Wafer-Scale Engine) for AI and high-performance computing, used to power SWE 1.5.
• GLM4.5: A model from the General Language Model series, known for its coding capabilities.
• Quen 3 coder: Another AI model, likely a fine-tuned version for coding tasks.
• Minimax: A highly capable AI model mentioned as performing significantly better than Cursor Composer and SWE 1.5.
• Kilo: Another AI model that is noted as being much better than the reviewed models.
• Diffusion LLM: A type of large language model that uses diffusion processes, mentioned as a faster alternative for tool calls.
• Tool Calls: A capability of LLMs to interact with external tools or APIs to perform specific actions.
• RL (Reinforcement Learning): A machine learning paradigm where an agent learns to make decisions by taking actions in an environment to maximize a reward.
• Fine-tuning: The process of adapting a pre-trained model to a specific task or dataset.
• Open Source: Software or models whose source code or weights are publicly available for use, modification, and distribution.
• VS Code: Visual Studio Code, a popular source-code editor. The reviewer suggests these companies might be forking or rebranding existing VS Code extensions or integrations.
• CLI (Command Line Interface): A text-based interface for interacting with a computer program.

Logical Connections Between Sections

The summary progresses logically from introducing the new models and their claims to a detailed performance analysis of each. The sponsorship section is placed between the introductions and the detailed testing, serving as a break and offering a practical AI tool for viewers. The criticisms are woven throughout the performance reviews and then consolidated in the "Key Arguments" section, highlighting the reviewer's core concerns about transparency, quality, and development practices. The technical terms are explained as they appear or are grouped for clarity.

Data, Research Findings, or Statistics

• Cursor Composer Speed Claim: Four times faster than similarly intelligent models.
• Cursor Composer Latency Claim: Completing most turns in under 30 seconds.
• SWE 1.5 Speed Claim: 950 tokens per second.
• Leaderboard Positions: Cursor Composer at 11th, SWE 1.5 at 19th.
• Company Valuation: Cursor is mentioned as a "$10 billion company."

Section Headings

• Key Concepts
• Model Performance and Criticisms
  • Cursor Composer
  • SWE 1.5 (Windsurf/Cognition)
• Sponsorship: Photogenius AI
• Methodologies and Frameworks
• Key Arguments and Perspectives
• Notable Quotes and Significant Statements
• Technical Terms and Concepts Explained
• Logical Connections Between Sections
• Data, Research Findings, or Statistics Mentioned

Synthesis/Conclusion

The video critically evaluates two new AI coding models, Cursor Composer and SWE 1.5, both claiming significant speed improvements for agentic coding and purportedly based on Mistral 4.6. The reviewer finds both models to be severely underperforming, generating "trash code" and failing on multiple basic tasks, despite their speed claims. A major point of contention is the lack of credit given to the underlying open-weight models, which the reviewer views as unethical and indicative of a lack of genuine innovation. The reviewer contrasts these commercial offerings with superior open-source models like Minimax and Kilo, arguing that companies should either leverage open models transparently or invest in developing their own from scratch. The sponsorship highlights Photogenius AI as a practical and affordable AI creation suite. Ultimately, the video strongly advises against using Cursor Composer and SWE 1.5 due to their poor quality and questionable development practices.