Reflection AI

Reflection AI is building open frontier models and a deployment stack for developers, enterprises and public-sector organizations. Its planned stack combines customizable models, software for running agents, and production infrastructure, supporting deployment on customers’ infrastructure or through partners. Reflection launched Asimov in 2025 for code comprehension: it brings together codebases, architecture documents and team discussions so engineers can understand systems and preserve shared knowledge.

Former DeepMind researchers Misha Laskin, CEO, and Ioannis Antonoglou, CTO, founded Reflection in 2024. Its engineering approach connects model training with agent applications. Asimov’s launch architecture paired multiple long-context retrieval agents with a short-context reasoning agent that combined their findings; that version used third-party models. In 2025, Reflection expanded from autonomous coding toward general agentic reasoning, announcing a mixture-of-experts training platform combining large-scale language modeling and reinforcement learning.

Reflection’s 2026 financing closed at a $25 billion pre-money valuation, confirmed by its CEO in April. Infrastructure commitments also grew: it signed a $1 billion compute-capacity agreement with Nebius in July 2026. These purchases support model development; as of July 2026, Reflection had released Asimov but had not launched a public foundation model.