The long-context frontier: own-model autonomous software engineering
HQ: San Francisco, CA | Founded: 2022 | Employees: Small, research-heavy (undisclosed) | Stage: Private; $1.5B (Aug 2024); 2026 secondary ~$1.1B | Website: https://magic.dev
Magic was founded in San Francisco in 2022 by Eric Steinberger (ex-Meta FAIR researcher with a poker-AI background) to build autonomous software engineers by owning the model stack. It released LTM-1 in June 2023 with a 5M-token context window - the defining long-context coding model of its era - and announced 100M-token context in August 2024 alongside a Google Cloud partnership for a custom TPU supercomputer and the $320M Series B (Eric Schmidt, Atlassian, Sequoia, Jane Street). As of 2026 it has raised $515M, markets Magic (coding assistant) and Magic Apply (autonomous job-application agent), and states it trains on thousands of NVIDIA GB200s. A claimed 2025 round with OpenAI/SoftBank interest could not be verified; secondary trading in 2026 marks ~$1.1-1.2B.
Magic's bet is that the model, not the IDE, is the unlocking layer of software engineering - so it trains its own frontier code models with giant context windows (5M to 100M tokens) on massive owned compute. The 2024 Series B ($320M at $1.5B, Eric Schmidt/Atlassian/Sequoia/Jane Street) proved the thesis's crowd. The 2026 reality is more sober: no new priced round, secondaries marking ~$1.1-1.2B (under the last price), and vanishing growth press while Cursor (~$60B), Codex, and Claude Code eat headlines. The bull case: long-context + own stack is the real moat in agentic coding; the bear case: a research-heavy company with no disclosed revenue facing well-capitalized distribution giants.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki