GPT-5.2 Pro Smashes Math Benchmark as DeepMind Predicts AGI by 2028
AI Benchmarks & Model Performance
GPT-5.2 Pro Sets New Record on FrontierMath Tier 4: OpenAI’s GPT-5.2 Pro achieved a 31% score on the FrontierMath Tier 4 benchmark, a significant leap from the previous high of 19%, as reported by Epoch AI Research.
AGI Developments & Predictions
DeepMind Chief AGI Scientist Forecasts 50% Chance of Minimal AGI by 2028: Shane Legg, Chief AGI Scientist at DeepMind, shared a prediction that there is a 50% probability of achieving Minimal AGI within the next four years.
New Models & Open-Source Releases
GLM-4.7-Flash-REAP Supports 200K Context Window on RTX 5060 Ti: The GLM-4.7-Flash-REAP model was tested on an RTX 5060 Ti 16GB GPU, achieving high performance with a 200K context window. Users noted LM Studio’s new feature for offloading model weights to CPU to handle larger contexts.
Sweep: Open-Weights 1.5B Model for Next-Edit Autocomplete: SweepAI released a 1.5B-parameter open-weights model optimized for predicting next code edits, outperforming larger models in speed and accuracy. The model includes a JetBrains plugin for integration.
AI Applications & Projects
Client-Side AI Plays Pokémon Red Using Qwen 2.5 1.5B: A developer built a 100% client-side AI that plays Pokémon Red using Qwen 2.5 1.5B via WebLLM and a neural network policy. The project is open-source and includes a live demo.
API & Service Updates
Devstral 2 Shifts to Paid API Access: MistralAI announced that Devstral 2 will transition to paid API access starting January 27, though free usage remains available under the Mistral Studio Experiment plan. A new release is teased for the following week.