Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,541 words
Add missing citations, update stale details, or suggest a clearer explanation.
Magic (legally Magic AI, Inc., often referenced by its primary domain magic.dev) is an American artificial intelligence company headquartered in San Francisco that develops large language models and autonomous agents for software engineering. Founded in 2022 by Eric Steinberger and Sebastian De Ro, the company has positioned itself as an attempt to build an "AI colleague" or "coworker" for code, rather than the autocomplete-style "copilot" products that dominated the early generative-AI coding market.[^1][^2]
Magic is best known for two technical bets that distinguish it from rivals such as [[cognition_ai]] (the maker of [[devin]]), [[cursor]], [[claude_code]], and [[github_copilot]]: an emphasis on ultra-long-context language models, culminating in an August 2024 research preview of LTM-2-mini, which the company described as having a 100 million-token context window, and the use of a custom long-context evaluation called HashHop designed to replace "needle in a haystack" tests.[^3][^4] The company has also built much of its identity around a small, secretive research team, a stated mission of "safely deploying AGI," and a strategy of training proprietary frontier models rather than fine-tuning third-party systems.[^5]
By the end of 2024, Magic had raised more than $465 million in disclosed funding from investors including Nat Friedman and Daniel Gross's NFDG, Alphabet's CapitalG, Eric Schmidt, Atlassian, Sequoia, Jane Street, and Elad Gil, and had announced a partnership with [[google_deepmind]]'s parent Alphabet's Google Cloud to construct two AI training supercomputers called Magic-G4 and Magic-G5.[^4][^6]
| Field | Value |
|---|---|
| Founded | 2022[^1] |
| Founders | Eric Steinberger (CEO); Sebastian De Ro (CTO)[^1][^7] |
| Headquarters | San Francisco, California[^2][^4] |
| Industry | Artificial intelligence; developer tools; foundation models[^4] |
| Status | Private; reportedly valued at approximately $1.5 billion in mid-2024 funding talks[^8] |
Eric Steinberger, an Austrian-born researcher, became interested in artificial intelligence in his early teens in Vienna and began collaborating with academic and industry researchers while still in secondary school, working with Johannes Heinrich, then a researcher at [[deepmind]], on reinforcement-learning projects.[^5] He went on to study computer science at the University of Cambridge and worked as a deep reinforcement learning researcher at Meta's Fundamental AI Research (FAIR) lab, where he collaborated with Noam Brown on game-theoretic and multi-agent learning problems.[^5][^7] In 2019 Steinberger also co-founded ClimateScience, a non-profit climate-education organisation, of which he served as chair.[^5]
Sebastian De Ro, Magic's co-founder and chief technology officer, comes from an industry rather than academic research background. He studied higher informatics at HTBLVA Spengergasse in Vienna between 2013 and 2018 and held engineering and leadership roles at the Austrian and German software companies Automic Software, twinformatics GmbH, and the business-process-management firm FireStart, where he rose to development lead, VP of technology, and finally chief technology officer before leaving to co-found Magic.[^9] Steinberger and De Ro met through ClimateScience, where De Ro volunteered, before partnering on Magic in 2022.[^9]
The company was originally pitched as a "coworker, not copilot" alternative to GitHub Copilot. In February 2023 Magic emerged from stealth with a $23 million Series A round led by Alphabet's growth-equity arm CapitalG, with participation from [[nat_friedman]], [[daniel_gross]], Elad Gil, and Amplify Partners, bringing total disclosed funding to approximately $28 million.[^1][^10] At the time of the Series A the company described itself as a six-person, distributed team building "an AI-powered pair programmer."[^1] Friedman's involvement was widely interpreted in the industry as a significant endorsement: he had run GitHub at the time of the original [[github_copilot]] launch in 2021 and so had unusual visibility into how AI-assisted coding tools were used at scale.[^1][^10]
Magic's central technical bet is that producing a useful autonomous software engineer requires language models capable of reading and reasoning over far larger code contexts than typical chat-tuned LLMs. The company has pursued this through a model family it calls LTM (for Long-Term Memory), branded as a "neural network architecture designed for giant context windows."[^2]
The first prototype, LTM-1, was disclosed in 2023 and was described as supporting a 5 million-token context window — a figure that at the time greatly exceeded the 8k-to-32k context windows offered by other commercial LLMs and was sufficient, the company said, to ingest most code repositories.[^2][^7]
On 29 August 2024 Magic announced its successor, LTM-2-mini, which it described as its first model with a 100 million-token context window. The company claimed this was equivalent to approximately 10 million lines of code or 750 novels.[^3] Magic emphasised efficiency rather than raw scale alone: it stated that the sequence-dimension algorithm used by LTM-2-mini was "roughly 1000x cheaper" per decoded token than the attention mechanism used by Meta's Llama 3.1 405B at the same context length, and that running Llama 3.1 405B at a 100 million-token context window would require "638 H100s per user just to store a single 100M token KV cache," whereas LTM-2-mini's memory footprint fit within a small fraction of a single H100 GPU's HBM per user.[^3]
Magic has also said it is training a larger production model in the LTM-2 family on the supercomputers it is building with Google Cloud, although as of mid-2026 it has not publicly released benchmark numbers or weights for that larger model.[^3][^4] In its 2024 announcement the company framed LTM-2-mini explicitly as a research preview: it stated that an at-scale LTM-2 production system would target dramatically improved reasoning over very long inputs, and that the company viewed both the model architecture and the training infrastructure as joint prerequisites for a useful coding agent.[^3]
Alongside LTM-2-mini, Magic introduced and open-sourced a new long-context evaluation called HashHop. The company argued that prior long-context tests, such as the popular "Needle in a Haystack" evaluation, leak semantic information that allows models to short-circuit retrieval. HashHop replaces those needles with random, incompressible hash strings; the model is asked to complete chains of hash-to-hash mappings, with the order of pairs shuffled to test position invariance and the chain length controlling how many "hops" of reasoning are required.[^3] Because the hashes contain no semantic redundancy, a model that scores well on HashHop must, Magic argued, store and retrieve from close to the maximum possible information density of its context window.[^3] Magic released a HashHop reference implementation on its blog and presented per-step accuracy curves for LTM-2-mini at varying context lengths, emphasising that the model retained competitive performance well beyond the points at which conventional attention-based systems became infeasible to deploy.[^3]
Magic disclosed in August 2024 that it had agreed a partnership with Google Cloud to build two purpose-built AI training systems:
Some press accounts of the partnership reported a peak aggregate performance figure of up to 160 exaflops across the two systems.[^11] Magic's own careers page describes the company as operating "thousands of GB200 GPUs."[^12]
The core product Magic has been working toward is an autonomous coding agent that the founders have repeatedly described as a "coworker" or "AI colleague" capable of writing, reviewing, debugging, and planning code changes across an entire codebase, rather than offering inline autocompletion.[^1][^7]
At its 2023 Series A, Magic said it was running an alpha programme with a small group of early users and intended to expand access as its training cluster and back-end stability improved.[^1][^2] At the time of its $320 million Series C in August 2024, TechCrunch reported that Magic's tools were still "not yet for sale," and that the company employed roughly two dozen people.[^4] As of the company's own public materials in 2025-2026, Magic continues to describe its products as research-stage and has not announced a general-availability commercial launch comparable to those of GitHub Copilot, Cursor, or Claude Code.[^12]
Beyond the model itself, public references from Magic's own materials and contemporaneous press have described a command-line interface that allows developers to "pair" with the model on a local codebase, and a cloud development environment for using the model and managing Magic accounts.[^2]
The "coworker, not copilot" framing has been a persistent feature of how Magic positions its product. In Magic's framing, an autonomous coding agent should be able to take a high-level request — implement a new feature, refactor a subsystem, fix a bug — and produce a complete, repository-aware solution that a human reviewer can accept, request changes to, or reject, in a workflow more analogous to that of a junior engineer submitting a pull request than to that of an inline autocomplete tool.[^1][^5] This contrasts with the autocomplete-first design of tools such as GitHub Copilot circa its 2021-2023 releases, although by 2024-2025 most major coding-AI products had moved toward more agentic, multi-step workflows themselves.[^4][^8]
Magic has raised at least four disclosed equity rounds, with publicly reported aggregates rising from roughly $28 million in early 2023 to about $465 million after its August 2024 round.[^1][^4]
| Round | Date | Amount | Lead / notable investors |
|---|---|---|---|
| Seed (cumulative pre-A) | through 2022 | approximately $5 million | undisclosed[^10] |
| Series A | 6 February 2023 | $23 million | CapitalG (lead); Nat Friedman, Daniel Gross, Elad Gil, Amplify Partners[^1][^10] |
| Series B | 16 February 2024 | $117 million | NFDG (Nat Friedman and Daniel Gross, lead); CapitalG, Elad Gil[^13] |
| Series C | 29 August 2024 | $320 million | Eric Schmidt, Jane Street, Sequoia, Atlassian; alongside continuing investors[^4][^6] |
Reuters reported in mid-2024 that Magic was in talks to raise additional capital at a valuation of approximately $1.5 billion — a more than threefold increase over an implied $500 million valuation reportedly attached to the February 2024 Series B — before the company's $320 million Series C closed in August.[^8][^4] Reuters and Crunchbase News also noted at the time that Magic "has no revenue and no product for sale," underlining how much of the company's valuation rested on its research roadmap rather than current commercial traction.[^8]
Magic's investor base went through a period of public uncertainty during 2025. In June 2025, Meta agreed to take a stake in Friedman and Gross's NFDG vehicle and to hire both partners to help run a new Meta AI division, an arrangement first reported by CNBC and Axios.[^14][^15] NFDG had led Magic's Series B and continued as an investor in the Series C; the change in the partners' day-to-day roles did not, as of the time of writing, produce any disclosed change in Magic's cap table or independent status.
Magic's most prominent disclosed partnership is the Google Cloud agreement announced concurrently with its August 2024 Series C, under which Magic is building the Magic-G4 and Magic-G5 supercomputers on Google Cloud infrastructure.[^4][^6][^11] The Magic-G5 system, based on NVIDIA Blackwell GB200 hardware, is intended to scale to tens of thousands of GPUs and to host the training of Magic's next-generation LTM-2 models.[^4][^6]
The company also disclosed a strategic relationship with [[atlassian]], which participated in the August 2024 round; press coverage of that round positioned Atlassian's involvement as relevant to potential deployment of Magic's coding agent inside Atlassian's Jira and developer-tooling ecosystem, though no specific product integration has been formally announced.[^4] CapitalG's involvement, dating back to the Series A, also gives Magic an indirect tie to Alphabet, although Magic's compute partnership with Google Cloud is described as commercial rather than equity-linked.[^4][^16]
Magic operates in one of the most heavily funded segments of the generative-AI industry. The company's most direct rival on the "autonomous software engineer" framing is [[cognition_ai]], whose Devin product was demonstrated publicly in March 2024 and which raised at a $2 billion valuation in mid-2024 — a higher valuation than Magic's own at the time despite Cognition being a younger company.[^8] In the autocomplete and IDE-integrated coding category, Magic's broader competitive set includes [[github_copilot]] (Microsoft), [[cursor]] (Anysphere), Codeium/Windsurf, Anthropic's [[claude_code]], and a range of agent-style coding tools released through 2024-2026.[^8]
Steinberger has publicly argued that, contrary to the assumption that value will accrue at the application layer, Magic's contrarian bet is that the largest economic and technical leverage will be at the model and "AGI" levels themselves, and that proprietary, frontier-scale models trained for code are the right substrate for an autonomous engineer.[^5] He has also argued that achieving useful "minimum viable AGI" in software engineering does not necessarily require capital expenditure on the order of $100 billion, making focused, well-funded specialist labs viable competitors to general-purpose frontier labs such as OpenAI, Anthropic, and [[google_deepmind]].[^5]
Magic remains an unusually secretive company by the standards of well-funded AI startups. It has published only a small number of technical posts (notably the LTM-1 announcement and the 2024 LTM-2-mini and HashHop write-up); it has not released model weights; it has not published full peer-reviewed evaluations of its LTM-2-mini system; and it has not, as of 2025-2026, announced a generally available paid product.[^2][^3][^12] These choices have drawn both criticism — that valuations are based largely on narrative rather than measurable product — and supportive comparisons to other small frontier-research labs that prioritise long research cycles over short-term shipping.[^8][^4]
Magic's competitive narrative also rests on the claim that ultra-long context is a category-defining advantage for code-focused models. While general-purpose frontier labs such as [[google_deepmind]], OpenAI, and Anthropic have steadily extended their own context windows through 2024-2025 — with publicly disclosed limits reaching into the millions of tokens — Magic's 100-million-token claim for LTM-2-mini remained, at the time of its disclosure, at least an order of magnitude beyond any comparable public commercial system.[^3][^8] Whether that gap translates into a sustained product advantage, once Magic ships a generally available agent, has been a recurring question in coverage of the AI-coding sector.[^4][^8]