Always Use Max Models in Cursor IDE

If you've been using Cursor IDE since its early days, you might have noticed a bizarre decline in quality around the time Claude 3.7 Sonnet was released. What used to be a reliable AI coding assistant suddenly became frustratingly obtuse - missing obvious references, forgetting about files you explicitly added to context, and generally acting like it had severe memory issues.

The common assumption was that Claude 3.7 Sonnet was somehow inferior to 3.5. After all, the timing matched perfectly - Cursor introduced some larger changes to its IDE around the same time, moving the agent into the chat window and apparently altering how context was handled.

But it's not Claude's fault at all.

Cursor's default behavior with standard models is to send only a pruned or summarized portion of your context. For example, when referencing a 1000-line script, the AI might only receive around 15 lines with an ellipsis indicating omitted content. Similarly, when opening a 4,000+ line file, Cursor will only send an outline and the region around your cursor to the model, not the entire file.

This context reduction has been confirmed by Cursor's team. They explained that they use a "smart context system" which selectively prunes and limits the context sent to the LLM to avoid performance issues. In their experience, sending full files (especially multiple large files) to the model causes a drop in quality or speed, so the app attempts to include only what it deems most relevant.

The consequence? Standard models ignore large portions of your context, even if you have the tokens available in theory. Some users discovered that enabling the "longer context" option in Cursor made no difference - the AI still received roughly the same limited context and failed to consider all the data. One developer found that Cursor was feeding only about 10,000 tokens of context to the model (about 1/100th of Gemini's potential 1M-token window), causing it to miss obvious details.

These limitations lead to frustrating experiences where the standard models provide wrong answers simply because not enough of the relevant code or documentation was provided.

The Solution: Max Models and the Return to Greatness

"Max" models aren't actually different models - they're the same Claude 3.7 Sonnet and Gemini 2.5 Pro you've been using, but with a critical difference: they bypass Cursor's heavy-handed context pruning. Both Claude 3.7 Sonnet MAX and Gemini 2.5 Pro MAX modes are designed to give the AI access to the entire long context that the underlying models support. In other words, Max models remove the artificial truncation or summarization that standard Cursor applies.

Impressive Context Windows, Actually Used

Claude 3.7 Sonnet MAX unlocks the full 200,000 token context window for your prompts, while Gemini 2.5 Pro MAX stretches all the way to a 1,000,000-token window. The Max models effectively turn off many of the restrictive filters, allowing the assistant to consider all the code and text you attach, rather than an abridged version.

The "Aha!" Moment for Developers

In practical terms, this makes a huge difference. Claude 3.7 Sonnet MAX can handle massive inputs and agent tool-chains that span hundreds of steps. In agent mode, Claude Max will process up to 750 lines of a file per read operation (far more than the few hundred lines standard mode might grab). Likewise, Gemini 2.5 Pro MAX lets you load entire modules or big documents into one session, something impossible with the standard 120k window in Cursor.

Getting Back the Original Cursor Magic

Early adopters report that with Max models, the AI can finally see all the functions and definitions relevant to a query, making its answers far more coherent and reliable for large-scale code understanding. Essentially, the Max variants restore the full context experience that early Cursor users fell in love with - it's like getting back the original Cursor experience, but with even more powerful underlying models.

Understanding the Business Decisions

Cursor has implicitly acknowledged these differences. The company's documentation and updates make it clear that the base models were constrained for practical reasons - primarily speed and cost. Indeed, Cursor's team admitted that they couldn't absorb the expense of huge context calls under normal subscriptions, which is why Max mode is a pay-per-use premium feature.

The Inevitable Monetization Strategy

Some in the community even speculated that Cursor intentionally nerfed the standard models' context to push heavy users toward the paid Max tiers. While Cursor frames it as offering a choice for "power users" who need maximum capability, the reality is that Max models are the only reliable way to utilize full context in Cursor.

Specific Use Cases Where Max Models Shine

Large Codebases and Legacy Systems

For developers working with monolithic applications or legacy codebases, Max models provide a comprehensive understanding that standard models simply can't match. When you need the AI to grasp intricate relationships between multiple files or modules, Max models can see the entire picture.

Framework-Heavy Development

If you're working with large frameworks like Laravel, React, or Angular, Max models can ingest the entire application structure, enabling the AI to provide relevant suggestions that account for framework-specific patterns and practices.

Working with Documentation and Specifications

When you need to reference extensive documentation, specifications, or requirements alongside your code, Max models can hold all of this information in context simultaneously, providing more accurate and contextually relevant assistance.

Max or Nothing

If you're doing complex stuff or working with large files, extensive codebases, or long documents in Cursor IDE, always opt for the "Max" model variants (Claude Sonnet Max or Gemini Pro Max). All evidence shows that the standard versions will truncate, summarize, or outright ignore much of your provided context beyond a certain point.

Worth Every Premium Credit

For serious Cursor users who demand accuracy and completeness in AI assistance, the Max models aren't just a premium option - they're essential for full functionality and context reliability. To get Cursor's AI to really "see" your whole project, Max mode is the only way to go. The difference isn't subtle - it's like switching from a frustrating, forgetful assistant to a brilliant collaborator who actually remembers everything you've shown them.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com