o3 mini high loses so much context all the time, even though it should have better context because of the thinking time. Also the code and problem solving is much worse.
o3 mini high is good for creating long step by step plans but claude sonnet is soo much better at implementing the plan and solving coding problems.
there’s no model that’s better than Claude sonnet in coding yet. o3-mini is good but still not good enough
Sonnet is since the launch on July 2024 untouched in coding and AGENTIC perfomence.
While other thinking models like R1, o3 mini and so on maybe are better for complex debugging they all struggle with Agentic perfomence like applying code and the biggest problem is that these models dont have any sense for Frontend design. I dont know how Anthropic achieved it but Sonnet creates stunning websites with the right prompt. My gut feeling also say they are impoving this model silently but i cant prove it. Also today i made new websites and i felt again that it has improved over the past days. All in all there is not a single competitor for Anthropic in daily automatic coding. And this makes me kinda sad because the market needs competition to push it to the limits.
I find Deepseek-r1 to be the best at planning personally
+1, sonnet3.5 is miles ahead on the agentic side fetching context etc.
100% agree. Would absolutely love to know what they have done differently to all the rest!
In my experience, o3 is better at “analyzing the big picture”. But it’s API is buggy and frequently fails to apply changes, and the model will be lazy and even outright gaslight you at times. Claude is the better coder, albeit with caveats that sometimes it goes off and does its own unrelated thing, doesn’t seem to see quite as much of the big picture, and has a tendency to “wallpaper over problems”.
Each have their own role, though in general, Claude is better. But if there’s a difficult problem, I recommend doing one round of o3 agentic at the top of the thread, tasked to analyze codebase about the problem, before handing off to Claude.
My observations exactly. o3 is very good at creating abstract outlines of things to be done, but it fails miserably at actually doing them. I will try your approach. Using o3 to create a road map, using sonnet for a step-by-step implementation.
I use o3-mini high because it’s still good at one-third the price of 3.5 Sonnet, even though there are often issues with applying changes or generating code. I use Sonnet 3.5 for solving more complex problems.
I feel if cursor team manage to build a Cascade Flow by windsurfer , we gonna be so god like. Regardless of models
Programming requires different models for different scenarios, making it difficult for one model to be suitable for all situations. For example, Claude is suitable for quickly creating overall project files and framework structures; o3-mini and r1 are suitable for writing complex logical reasoning algorithms; o1-mini is suitable for solving problems with logical complexity between Claude and o3-mini, but o1-mini has much stronger text capabilities than o3-mini; deepseek-v3 can also solve many ordinary programming problems, but it is a bit slow; gpt4o-mini is only suitable for solving very simple problems, but it is very fast; in large projects, local generation using Tab and Ctrl+K is more stable; currently, we are still exploring how to make good use of Composer in large projects. For a given problem, we can try to solve it using different models to see which model can handle it.
This is a gameplay problem regarding an item in a game:
The Heart Shield can be synthesized from two mirrors of the same star level to create a higher star level Heart Shield, with a maximum synthesis of a nine-star Heart Shield. The newly synthesized Heart Shield will inherit the attributes of the two participating mirrors.
The rules are as follows:
1 star (poison resistance) + 1 star (poison attack) = 2 stars (poison resistance, poison attack)
1 star (poison resistance) + 1 star (poison resistance) = 2 stars (poison resistance, poison resistance)
2 stars (poison resistance, poison attack) + 2 stars (poison resistance, poison resistance) = 3 stars (poison resistance, poison attack, poison resistance)
1 star (poison resistance) + 1 star (poison resistance) = 2 stars (poison resistance + random attribute)
Each star level of the mirror can contain a different number of attributes:
1-star Heart Shield has 1 attribute
2-star Heart Shield has 2 attributes
3-star Heart Shield has 3 attributes
4-star Heart Shield has 4 attributes
5-star Heart Shield has 5 attributes
6-star Heart Shield has 5 attributes
7-star Heart Shield has 6 attributes
8-star Heart Shield has 6 attributes
9-star Heart Shield has 7 attributes
From the above rules, it can be seen that the attributes of the Heart Shield synthesized are generated based on the attributes of the materials used. However, if the target mirror can contain more attributes than the union of the attributes of the two material mirrors, the fewer attributes will appear randomly. Therefore, to stably obtain the desired attributes in the resulting mirror, this situation should be avoided, and only mirrors of the same star level can be merged.
Now I have enough 1-star material mirrors with different attributes. Please write Python code that takes the target attributes and target star level as input and generates the synthesis steps.
，我英语不好,下面是中文：I noticed there are several different Sonnet models - whats the difference between them? should they all be enabled?
I’m suspecting that Sonnet is iteratively updated, but the updates are not announced.
100%
