Top AI Picks Of 2025

One of the things I did with the family over Christmas/New Year was for each of us to lay out the top things we experienced in 2025 - top movie, top album etc.

I thought it would be interesting for me to do something similar from an AI centric viewpoint, so here goes:

➡️ Top Foundational Model

I think if this question was asked at the beginning of November I would have punted for Claude Sonnet 4.5, but then with the release of Gemini 3 mid November, and then the subsequent release of Claude Opus 4.5 at the end of November....and Chat GPT 5.2 in the second week of December, this makes picking just one very difficult as all models have taken a huge step forward.

Given that a tie would be a cop out, I’m going to give it to Gemini 3 as I find myself using Gemini 3 Pro and Gemini 3 Flash API way more than I used to.

Gemini 3 👉 https://blog.google/products/gemini/gemini-3/

➡️ Top SLM

This one is easier, I’ve done a number of projects now with OpenAI’s local GPT-OSS models and they are very capable, particularly for well defined local use cases.

The 20B variant also runs on fairly commodity machines as GPT-OSS SLM‘s are distributed with a native MXFP4 4-bit weight format to enable efficient inference on smaller hardware. This low-bit format significantly reduces memory/compute requirements, allowing it to punch significantly above its weight class while remaining efficient.

GPT-OSS 👉 https://openai.com/index/introducing-gpt-oss/

➡️ Top AI Lib

There is a lot to choose from here but I am going with Google’s LangExtract.

LangExtract was released as an open-source Python library to solve the “black box” problem of LLM data extraction. It allows developers to extract structured data (i.e. JSON) from unstructured text (i.e. PDFs, medical records, legal contracts etc) while maintaining precise source grounding. Unlike many common extraction approaches, every piece of data LangExtract pulls out is linked back to the exact character span in the original document.

Although it is optimized for the Gemini family, it is model-agnostic and supports OpenAI models and local models via Ollama.

👉 https://developers.googleblog.com/introducing-langextract-a-gemini-powered-information-extraction-library/

➡️ Top AI Pattern

In 2025, from my perspective of working with companies, this was still Retrieval Augmented Generation (RAG), although towards mid to end of year this started to shift towards AIAgents. Many enterprises deployed things such as Microsoft Copilot/Amazon Q or built their own RAG proof of concept.

However, because unstructured data is messy, fluid, and frequently conflicting, RAG deployments, particularly at scale, that lack strict architectural discipline tend to break, produce inconsistent results, and incur high costs. This is why we’re now seeing hybrid patterns emerge i.e. RAG for grounding, agents for orchestration, and tighter data contracts between the two.