AIHow AI Works

Tag: Reasoning

OpenAI Launches GPT-6 Astra With Record Benchmarks

OpenAI announced GPT-6 Astra on September 3, 2026, claiming 97.6% on FrontierMath Tier 4 v2 and 100% on ExploitBench. The ARC-AGI-3 result comes with a caveat.

GPT-5.6 Ships as Three Models That Differ Only in Price

OpenAI's Sol, Terra and Luna went generally available July 9, 2026. All three share one context window, one output cap and one knowledge cutoff.

Thinking Machines Releases Inkling, Its First Open-Weights Model

Thinking Machines Lab ships Inkling, a 975B-parameter multimodal MoE under Apache 2.0 — the leading U.S. open-weights model at release, per Artificial Analysis.

Training vs Inference: Two Economies, Two Opposite Bottlenecks

Training is network-bound and paid once. Serving is memory-bound and paid forever. The crossover point, and why the 2024 chip-demand call was backwards.

Google Gemma 4: The Next Frontier of Open Models for AI Agents

Google introduces Gemma 4, a new family of open models optimized for complex reasoning, autonomous agents, and tool use with up to 256K context window.

Gemini 3.1 Pro: A Smarter Model for Complex Tasks

Google announces Gemini 3.1 Pro, featuring a 77.1% ARC-AGI-2 score and advanced reasoning for agentic workflows and complex system synthesis.

Tencent HY 2.0: MoE Model with 73.4 IMO Score

Tencent releases HY 2.0 foundation model with MoE architecture, 256K context, and major gains in reasoning, coding, and instruction following capabilities.

Step Audio R1: First Audio Reasoning Model

Step Audio R1 is the first audio model to unlock Chain-of-Thought reasoning, solving inverted scaling and surpassing Gemini 2.5 Pro in complex audio tasks.

DeepSeek-V3.2: GPT-5 Level Reasoning & Agent AI

DeepSeek releases V3.2 and V3.2-Speciale models with GPT-5 level performance, gold-medal reasoning capabilities, and thinking in tool-use for agents.

Qwen3-Max-Thinking: Perfect Reasoning Scores

Alibaba's Qwen3-Max-Thinking achieves 100% on AIME 2025 and HMMT, matching OpenAI's top model on reasoning benchmarks while emphasizing step-by-step solutions.