How To Optimize Your Usage: The Best AI Models to Use, version 2

How To
1

Previous version of the guide:

Overview

Auto

Great for quickly digesting codebases, light analysis, and simple edits. Use it whenever possible — it’s fast, reliable, and free. Seriously, there’s no reason not to default to Auto for routine tasks.

o4-mini

When Auto falls short or doesn’t quite get what you’re asking for, switch to o4-mini. It’s not great with long context, but you can count on it for straightforward code edits. Bonus: it’s about 3× cheaper than Gemini 2.5 Pro or Claude 4.

Claude 4

I’m not a huge fan of how proactive Claude can be. And I really don’t see the point in paying extra for the “Thinking” variant.
However, Claude 4 is a very good bloodhound when you need to dig into a problem or you’re too lazy to think for a prompt.

Gemini 2.5 Pro

The former king for complex multi-step tasks. Had persistent edit_tool issues in Cursor. It’s holistic approach to problem-solving makes it one of the most well-rounded models — right behind Grok 4 and o3-Pro.

o3-Pro

Think of this like a precision tool. Try it in Manual mode when you want deep, one-shot analysis over anything that you put in its context. Just remember — the minimum cost per request is $0.80. Are you sure the task is worth that?

Grok 4

The new champion. Cursor’s team is actively optimizing its integration — they even visited the XAI office to get it dialed in. Grok 4 clearly approaches problems differently. It does things that Gemini 2.5 Pro can’t.

Optimal Strategy

Budget-Friendly (Pro users)

  • Auto for everything until it breaks.
  • Gemini 2.5 Pro for complex tasks.
  • Grok 4 for one-shot analysis in Manual mode.

Balanced Strategy (Pro+ users)

  • Auto and o4-mini for minor edits.
  • Claude 4 for medium tasks, lazy prompts or as QA-engineer without code editing
  • Gemini 2.5 Pro for in-depth engineering tasks.
  • Grok 4 when you want to GROK THE REAL PROBLEM.

Premium (Ultra users)

  • Claude 4 or Gemini 2.5 Pro for quick edits and analysis.
  • Grok 4 MAX for everything else.

Turn on bell in Watching to get updates about the guide!

15 Likes
2

I released the guide just in time :joy:
2025-07-11_22-05-37

3

I often see that there are people who are glued to Claude and that’s it. If you have your own usage strategy, it would be interesting to read.

2 Likes
4

UPD: Added a connection of usage strategies and Cursor plans

5

I can also recommend you to use my Agent Compass.

If you think that this are ineffective, I will be glad to hear feedback on Github or in Compass topic on this forum.

6

Judging by the difference in the display of edit_tool, I have already come across Claude inside Auto several times. However, I do not know which version.

7

I’ve barely managed to get anything done with Grok 4 over the past couple of days due to how horribly it behaves in Cursor. At this point, I’d say it’s only usable for one-shot tasks and one-shot analysis.

1 Like
8
9

Agent Compass already has 12 stars! :scream:
Thanks so much for the support!

I’ll try my best to release Agent Enforcer for Python, C# .NET, and Android Kotlin within the next three days.

10
11

UPD: Claude as QA-AI in balanced strategy

12

Thanks for the guide I’ll try that.

btw, is Sonnet 3.5 one of the free models? It used to be great before 3.7 came out

13

Previously, Gemini 2.5 Flash and GPT-4.1 were free. Now - just use Auto

2 Likes
14

[MOD - Condor - changed category of link]

15

hi bro, thx for the usage. I agree with that gemini2.5

I think its action mode is just “planning next move” and output.Just like that and no more even 1 step. It dosent like claude which can call many tools very flexibly.
so how colud it become the in-depth engineering tasks for Pro+ users?
In this situation, I found that my usage scenarios for it became very limited, only when discussing software architecture or needing to explain problems in words. But in this case, I don’t think the revenue has increased much compared to aistudio on the web. So I would like to ask for your guidance on how to use it. Did I miss anything or have any techniques?
Because i really love gemini2.5 and want use it

16

Gemini is smarter than Claude. Although some Claude Code users have told me that Claude is dumber in Cursor than in Claude Code.

When working with Gemini, just follow the process of its operation. Try to stop and/or restart the task. It’s usually annoying, but it’s critical. If you can’t continue the dialogue at all, then switch to Claude, o4-mini or switch to a new chat.

1 Like
17
18

Hi,

You like to improve this, you better add what drains the tokens too. Like images and rules and other docs, long files, lots of nested imports, reusing same chat session and probably other stuff.

Starting a new chat for each prompt will save a lot of tokens, and at the end it is all about the number of tokens.

19

If the Agent has not done all the work that was supposed to be done, then it is worth continuing in the same chat to preserve the context. This way you will definitely not miss the context and will not waste time and tokens on re-searching the context.

1 Like
20

this needs to be addressed. a task must be well defined and thought through and well prompted. I speak from my own experiences. I used to get ideas along the way and continued on the same chat session and the conversation grew beyond recognition and token usage went through the roof. It means my tasks were not scoped correctly.
A well scoped task must finish within one prompt and if bugs occur, they can be fixed in the next chat, but if it is derailed that is another matter.

1 Like