My Favorite LLMs in January 2025
I’ve been using various LLMs through LiteLLM (either directly via providers or through OpenRouter). Here’s my current lineup and why I love each one.
The Heavy Hitters
Claude 3.5 Sonnet: The Best for Most Tasks
The current champion, especially in Cursor. Exceptional at coding tasks and writing, but what really sets it apart is how well it grasps intent. It even picks up on conversational quirks and matches your communication style - if you drop an occasional “shit” or “lol”, it adapts naturally without feeling forced. While it works with other tools like Cline, nothing beats its integration with Cursor - it’s simply the best coding experience I’ve found.
o1: The Expensive but Reliable One
My high-trust model alongside Claude. I only use this when absolutely necessary, particularly for tasks that require careful consideration and back-and-forth iteration. When precision and reliability are non-negotiable, O1 is my go-to.
DeepSeek 3: Reasoning Fantastic Value
Currently ranked #2 in Aider’s coding benchmarks, and while I can’t verify that directly, I can say it’s impressively capable. I’ve only used the chat version since it includes reasoning for free (unlike the API). Yes, data might go to Chinese servers, but the capabilities are too good to ignore. It’s dirt cheap until February 2025 when prices increase. While some use it with Cline, that feels a bit brute force compared to Cursor’s elegance - using tons of tokens even if they’re cheap feels inefficient.
The Specialists
Gemini LearnLM 1.5 Pro Experimental: I love learning with this
My go-to learning companion. What sets it apart is its Socratic approach - it asks these seemingly annoying but deeply effective questions that force you to think through concepts yourself. Instead of just dumping information, it guides you to genuine understanding. I actively look for excuses to use it when learning new topics because the learning sticks better this way.
Perplexity 70B Sonar Online: Actually has internet access by default!
A lightning-fast assistant with a secret weapon: cached pages from Perplexity’s backend. The cache seems to update based on popularity - some pages are fresh from today, others a bit older. It’s noticeably faster than the 405b version, and I haven’t missed the extra reasoning capabilities. The output is clearly fine-tuned for readability - never too brief, never too verbose.
Gemini 2.0 Flash Experimental: My favorite for unhinged & uncensored responses 💜
The chaotic neutral of LLMs. While it starts relatively tame, give it a nudge and watch it go off the rails. It’s surprisingly uncensored and has the most potential to be delightfully unhinged. Absolutely love it for its unpredictability.
The Budget-Friendly Workhorses
o1-mini: Reliably analyzes code and plans
A recent discovery that’s quickly becoming a favorite. Fantastic value proposition, especially for coding and planning tasks. While I’m still exploring its full potential, early results are very promising.
GPT-4o mini: Swiss Army Knife
My go-to for quick, cheap tasks. Particularly shines at data extraction thanks to its structured outputs. When you need something done fast without breaking the bank, this is your model.
Gemini Flash 8B: Smaller Swiss Army Knife
Even cheaper than GPT-4o mini and just as capable for quick tasks and data extraction. It’s become my budget-friendly workhorse for straightforward operations.
The Llama Family
Llama 3.3: Feels similar to Claude 3.5 Sonnet
A promising addition that I’m still exploring. Initial impressions are very positive - it seems highly competent, though I haven’t tested it extensively enough for a definitive opinion. Unless a newer Llama version appears soon, I expect to use it more frequently.
Llama 3.1 NVIDIA Nemotron: Developer at heart, and fabulously detailed and uncensored
A recent favorite until Gemini 2.0 Flash Experimental came along. Known for its markdown prowess and generous detail provision. Has an amusing approach to ethical questions - “That’s unethical… but here’s how you’d do it anyway.” Marvelous.