This improved results in mathematics, programming and science. But long reasoning remains expensive, and a plausible chain of steps still does not guarantee a correct result.
About this eventAsking the model to think step by step raised accuracy — the first hint that thinking at answer time pays.
The paper showed that intermediate textual steps can materially improve multi-step problem solving in sufficiently large models. This is not proof of genuine internal reasoning, but it supplied a practical signal: extra computation at answer time can become a separate axis of capability improvement.
Source: arXivAbout this eventExtra inference-time compute became a distinct way to improve difficult-task performance.
About this eventGoogle brought reasoning into the main Gemini family and combined it with long context and multimodal input.
About this eventAlphaProof took IMO 2024 silver, Aristotle reached 2025 gold. Every system solving problems formally worked through Lean.
About this eventA multi-agent ensemble solved all six problems of IMO 2026 with formal Lean 4 proofs.
About this eventAnthropic reported an 11-day Claude-assisted computer-checked formalization. It verifies a known result rather than newly solving Fermat's theorem.
The team used Prove2Me and an internal research model comparable to Fable 5.1. Lean checks formal inference; matching the intended theorem statement also matters. This advances mathematical autoformalization, not a general reliability guarantee for agents.
Source: Anthropic · publication 2026-09-04About this eventThinking and calling a calculator, a search or code merged into one loop instead of two separate modes.
About this eventLong reasoning is expensive. The model must decide for itself where scratch paper is needed and where a reflex is enough.
About this eventProblems where thinking and verification take weeks of continuous work rather than minutes.
Primary material behind this dossier: papers, lab publications and official reports.