Frontier models built by scaling reinforcement learning — from autonomous coding to superintelligence
HQ: Brooklyn, NY, United States | Founded: 2024 | Employees: ~100-200 (est.) | Stage: Late Stage VC ($27.5B valuation, March 2026) | Website: https://reflection.ai
Reflection AI was founded in 2024 by Misha Laskin (CEO) and Ioannis Antonoglou, both formerly of Google DeepMind. Laskin led reinforcement-learning work including Gemini's post-training and RLHF efforts; Antonoglou was an early engineer on AlphaGo and AlphaZero — the systems that pioneered superhuman agentic behavior through self-play RL. Their founding thesis: the next leap in AI value comes not from bigger language models alone but from scaling RL to create agents that act autonomously over long horizons — and coding is the ideal first domain because it is verifiable, economically massive, and generates perfect reward signals.
The company emerged from stealth in March 2025 with $130M in seed and Series A funding (co-led by CRV and Lightspeed) at a ~$545M valuation, launching Asimov, a code research agent for engineering teams. In October 2025 it raised $2B at an $8B valuation led by Nvidia — one of the largest rounds ever for a lab with early products — with participation from Sequoia, Lightspeed, Eric Schmidt, and Citi among others. By March 2026, Sacra reported a $2.5B round at a $27.5B valuation.
Reflection's roadmap is explicit: prove RL-scaled agents in coding, then extend toward general agentic systems — ultimately pursuing superintelligence via the same recipe that took Go from human-level to superhuman. It is among a new class of 'second-generation' frontier labs (alongside Safe Superintelligence, Thinking Machines) betting that post-training RL, not just pretraining scale, is where the frontier moves next.
Reflection AI is the boldest RL thesis in the post-ChatGPT era: the founders argue LLMs alone plateau, and the 'AlphaGo moment' for code and agents comes from scaling reinforcement learning on top of them. The market has priced that thesis aggressively — $130M at $545M (March 2025), $2B at $8B led by Nvidia (October 2025), then $2.5B at $27.5B (March 2026) — a 50x valuation jump in one year for a lab still shipping early products. The bull case: Laskin and Antonoglou come from the only team that ever solved agentic behavior with RL at superhuman levels (AlphaGo/AlphaZero), and autonomous coding is the highest-value first market. The bear case: it is a capital-intensive race against frontier labs (OpenAI, Anthropic, Google) that have vastly more compute, data, and distribution.
Category: AI & Machine Learning Startups
Browse: Home | Blog | About | All Startups A-Z | AI & Machine Learning Startups | View full profile on StartupWiki