Magic is a San Francisco AI lab building frontier code models with ultra-long context windows aimed at automating software engineering and AI research. Founded by Eric Steinberger and Sebastian De Ro, the company has raised roughly $515M from Sequoia, CapitalG, Nat Friedman, Daniel Gross, and Jane Street to pursue a direct path toward safe AGI through code intelligence. Magic is best known for its claimed 100-million-token context window, a research bet that an AI which can read entire codebases in a single prompt will out-reason agents stitched together with retrieval.
What Magic does
Magic trains frontier-scale code models using large-scale pre-training, domain-specific reinforcement learning, and inference-time compute scaling on a Google Cloud GPU cluster of thousands of NVIDIA GB200s. Its public artifact, LTM-2-mini, demonstrated reasoning over 100M-token contexts on a benchmark called HashHop, designed to test true long-range recall. The product roadmap points at fully automated software engineers and AI researchers rather than IDE plugins.
Who it's for
Magic is a research-led, enterprise-only company with no public self-serve product as of 2026.
- Frontier labs and AI researchers wanting to study long-context reasoning
- Enterprise R&D teams evaluating bespoke code automation pilots
- Investors and AGI watchers tracking the long-context thesis
How Magic compares
Magic competes with frontier model labs (Anthropic, OpenAI, Google DeepMind) and code-specialist labs (Cursor, Cognition, Poolside). Its differentiator is the long-context bet: where Cursor and Cognition Devin focus on agent harnesses and tooling, Magic argues that scaling raw context to 100M tokens removes the need for fragile retrieval. Versus Poolside, Magic leans more on inference-time compute and less on synthetic execution feedback.
Pricing and access
Magic has no public pricing or self-serve product. Access is enterprise-only through direct conversations and selective partnerships.
Why it matters in 2026
Long-context coding is one of the live debates of the AGI race. If Magic can ship a usable 100M-token model, it changes how teams build with AI: no chunking, no RAG glue, the whole repo in the prompt. With Sequoia and Alphabet behind it and a GB200 cluster online, Magic is one of the few independent labs still credibly chasing frontier code intelligence.