Reasoning Models
Reasoning models are a new class of AI that "thinks" before answering: spending extra compute at inference time, building chains of thought and solving problems ordinary LLMs cannot — math, code, science. They kicked off the test-time-compute race. Here is all our coverage of reasoning models: releases, benchmarks (GPQA, ARC-AGI), limits and cost.
Latent Space (swyx)Anthropic reveals how it builds Claude with Claude: the brain separate from the hands
AWS Machine Learning BlogBayer Launches Multi-Agent Engine on AgentCore: 80% of Questions Handled Without Engineers
Hugging Face BlogLiquid AI releases LFM2.5-24B-A2B: an agent model with 2B active parameters
arXiv cs.AIAgentOrch: Pre-Planned AI Agents Now Trail Step-by-Step Execution by Just 8.5 Points
NVIDIA AI BlogNVIDIA Releases 120B Nemotron 3.5 Super VL and Bolt Models for Local Agents
AWS Machine Learning BlogAWS Launches Multi-Turn RL in SageMaker: gpt-oss-20b Climbs From 15% to 69% for $339
Cognition BlogCognition releases FrontierCode 2.0: the best score is just 45.5%
Hugging Face BlogReflection Releases Beam: An Open 501B-Parameter MoE Model Under Apache 2.0
NVIDIA Releases 120B Nemotron 3.5 Super VL and Bolt Models for Local Agents
NVIDIA opened Nemotron 3.5 Super VL, a 120B-parameter model with a 5x speed gain, plus two Bolt models. Bolt 74B-A9B fits in 24 GB thanks to 2-bit quantization.
AWS Launches Multi-Turn RL in SageMaker: gpt-oss-20b Climbs From 15% to 69% for $339
SageMaker AI now fine-tunes agents on a single reward for an entire chain of steps. On Wordle, gpt-oss-20b rose from 61% to 84% of words solved; training tokens…
Cognition releases FrontierCode 2.0: the best score is just 45.5%
Cognition updated its coding-agent benchmark with larger tasks and stricter criteria. Claude Opus 5.5 scored 45.5%, while Kimi K3 managed only 17.7%, down from…
MLE-World predicts the outcome of an ML experiment without running it: 77.9% accuracy
OpenAI releases GPT-6.1 Terra: $1/$5 per million tokens and a focus on computer use
Thinking-in-Bets: Agents Lock In Claims Before Criticism and Sycophancy Drops
ThinkBit shrinks Qwen3-14B to 3.74 GB and still scores 94.8 on MATH-500
Shan Yang breaks down RL gains: 82% comes from fitting answers into the token budget
A comparison of 20 open RL models with their bases finds that low-budget pass@1 gains come from finishing answers. RL adds new solved problems by at most 2.1…
Anthropic Rewrote Its Take-Home Test Three Times: Claude Opus 4.5 Caught Up With the Best Engineers
1000+ people took the performance engineering take-home. Opus 4 beat most of them in 4 hours; Opus 4.5 matched the best in 2. Now the test is being rewritten a…
Maia 300: How Microsoft Is Cutting Its Dependence on Nvidia GPUs With Its Own AI Chip
Independent Critic Lifts Coding-Agent Accuracy by 5.9–9.5 pp
Meta releases Muse Wonder and Wonder Full: intelligence index rises from 48 to 56
Russia Already Has 18 AI Code Agents, but 'Russian' Means Three Different Things
University of Pittsburgh Looks for a 3D Map Inside a VLM: How a Model Sees a Room
Haoming Wang and Wei Gao test whether vision-language models store a room's objects as a map of adjacency relations or just as a bag of colors and shapes.
Apple and CMU Speed Up AI Agents 10x by Replacing Reasoning with Selection
Apple and Carnegie Mellon researchers propose dropping free-form "thinking" from ReAct agents and replacing it with a choice among options. The claimed speedup is…