I haven’t been able to compile what has happened in the AI World since October 2025 and a lot has happened. Here are some highlights from the last quarter of 2025..
OpenAI
ChatGPT Atlas
OpenAI has released a browser with an intrinsic link to ChatGPT. It uses one’s OpenAI account and it can use information on the Internet. This seems to be an attempt to get a piece of the Internet search engine pie.
GPT 5.1
OpenAI has released a new version of GPT with version number 5.1, which they declare to have a smarter and more conversational ChatBot. GPT‑5.1 Instant is the most-used model, whereas GPT 5.1 Thinking is the “reasoning” model which can do deep and iterative reasoning to get better results with more time allocated. GPT 5.1 adapts its working time so that it spends less time on simple tasks, but ends up spending more time on more difficult tasks.
GPT 5.1 Codex Max
OpenAI has released a new version of their coding frontier model. It is up to 15% faster and more efficient in token use as compared to OpenAI’s previous coding model GPT 5.1 Codex.
GPT 5.2
A short time after 5.1, OpenAI released the next flagship model, GPT 5.2. It has done much better on most benchmarks and is also better in coding.
ChatGPT Images 1.5
Right after GPT 5.1, OpenAI released ChatGPT Images 1.5. This release improved image generation speed, but especially gave the user more control over editing images.
GPT 5.2 Codex
OpenAI has released a new version of their coding frontier model. It is optimised for agentic coding and has stronger cybersecurity features as compared to earlier coding models. It can also generate better code as the session continues, since context is now larger. Design mocks can be translated into working prototypes.
DeepSeek
Sparse Attention
Sparse attention is a mechanism that can be used to optimise performance of transformers, by reducing the number of simultaneous “attention” operations a model will have to perform. DeepSeek seems to have solved the issue by implementing the sparse attention mechanism, thereby achieving the same performance with less cost.
v 3.2
DeepSeek has released two versions for their foundation model. Deepseek v. 3.2 is at GPT 5-level performance, according to the company and has gone live in December 2025. Deepeek v. 3.2 Speciale does more “thinking” at the expense of response time. These models use the sparse attention technique described in the previous paragraph.
Anthropic
Claude 4.5 Haiku
Anthropic has released their compact model dubbed Claude 4.5 Haiku, which has a similar performance with larger models.
Google
C2S-Scale 27B
Google Deepmind has produced a new specialised AI model named C2S-Scale 27B, based on their own Gemma model. Result of a collaboration with Yale University, the model has been used to suggest a new therapy for cancer, which is now being clinically and experimentally tested.
Veo 3.1
Google released a new text to video model which has impressive results, such as producing compound backgrounds, people, and objects into a coherent scene, including audio, receiving the first and final frames and filling in the gap and finally extending short clips with new frames.
Sima 2
Google Deepmind has released Sima 2, which is an Agent that plays, reasons and learns in virtual 3D worlds. It is powered by Gemini and can play video games along with a human providing instructions.
WeatherNext 2
Google Deepmind has released a new version of their weather prediction model, WeatherNext 2. This version is much faster than the previous one and can also create higher-resolution weather forecasts. Google will be integrating the model functionality to several of it’d own products.
Gemini 3 Pro Image (Nano Banana Pro)
In this update for their image generation model, Google improved many things, but especially text handling, generation of infographics and other visual tools and precise control over generated images.
I tried it with a no-so-complex prompt and got a quite realistic photo-like image. For some reason (?) the protagonist looks like Ben Affleck. This was done using the Fast mode and processing took less than 10 seconds.
Gemini 3 DeepThink
Gemini 3 has a new mode called Deep Think. It’s only available for Gemini Ultra subscribers ($250/month) and uses advanced parallel reasoning. It got quite good results in most benchmarks.
Gemini 3 Flash
Google released a new version of their smallest model, Gemini 3 Flash and it turned out to be a very fast model. It also seems to perform really good against similar small models.
Apple
Siri with Gemini
Apple announced that they have reached an agreement with Google to use their Gemini model to power the next version of Siri, Apple’s Assistant software. It is also understood that Gemini may not be the only model to be utilised.
Mistral
Devstral 2
Mistral released a new version of their coding model, Devstral 2. This was considered to be the first reasonable model coming out of a European company.