About
Induction Labs is a San Francisco research lab building foundation models that learn by observing the world, including computer-use behavior and internet video. It has not publicly named paying customers; its disclosed direction supports developers and teams building intelligent agents and computer-interacting products, differentiated by models that learn directly from video and Photon-1’s reported lower training and serving costs.
Market
Induction Labs competes in the frontier foundation-model and computer-use-agent market, but positions itself as a research-first model company rather than a conventional browser-automation platform. Its differentiation is the use of large-scale computer-demonstration video and latent-space world modeling to let models imagine future states before acting, with the company reporting lower training compute and serving cost than a production LLM and transfer to games and physical simulation. OpenAI and Google emphasize general-purpose GUI agents exposed through products or APIs, while Browser Use, Simular, and Browserbase focus more directly on deployable browser/computer-agent software and infrastructure.
Public materials do not disclose paying customers; Induction Labs is a small, two-person San Francisco research lab. Its best-supported prospective customer profile is frontier-AI labs, AI platform developers, and enterprise automation teams that need general-purpose computer-use agents or models capable of learning from demonstrations rather than relying only on manually specified APIs.
At a Glance
Problem
Induction Labs is addressing the bottleneck in building AI that can learn and act in open-ended environments. Today’s computer-use agents often depend on labeled action trajectories or laborious integrations, while general-purpose models must reason over pixels and navigate fragmented, legacy software. That limits reliable end-to-end automation and makes training expensive. Induction Labs’ broader thesis is that scalable intelligence should learn from its own observations, update as it goes, and develop knowledge without human tutors.
The clearest near-term use case is autonomous computer use: an agent that can operate desktop applications such as VS Code, Gmail, and ChatGPT, taking over repetitive digital workflows without requiring a custom API for every tool. Induction Labs reports that its Photon-1 model outperformed Gemini 3.1 Flash-Lite on an internal computer-use benchmark with 30 times less pretraining compute and roughly three times lower inference cost, directly targeting the capability and unit-economics constraints of digital labor automation.
Product / Service
Induction Labs is currently a research lab, and its initial product is Photon-1, a sparse 106B-A5B mixture-of-experts “imagination model.” The model was pretrained on approximately 575 million frames—equivalent to 18 years of computer-screen video—without action labels. It compresses visual states into latent tokens, predicts the next state of a computer environment, and is then fine-tuned on fewer than 35,000 labeled computer-use trajectories to emit the keyboard and mouse actions needed to reach that imagined state.
The intended delivery model appears to be foundation-model infrastructure for autonomous agents and workflow automation rather than a generally available end-user application. Photon-1 can simulate desktop sessions and, after fine-tuning, play checkers, model billiard physics, and use a ChatGPT-like system in a human-style loop. However, as of August 1, 2026, independent reporting indicates that there are no public weights, API, or license, so the immediate benefit is a research demonstration of a potentially cheaper, more data-efficient route to computer-using agents rather than a deployable commercial service.
Market
Induction Labs sits at the intersection of foundation models, computer-use agents, agentic coworkers, and world models. Its closest benchmark comparator is Google’s Gemini 3.1 Flash-Lite, while the broader computer-use market includes OpenAI’s ChatGPT Agent, Anthropic’s Claude and Claude for Chrome, Google’s Project Mariner, Manus, Context, UI-TARS, Qwen-VL, OpenCUA, and Simular S2. Induction Labs’ differentiation is its attempt to learn an implicit policy from unlabeled observational video by predicting future states, rather than relying primarily on action-labeled demonstrations or traditional browser automation.
The company’s public traction is research rather than broad commercial deployment. It was founded in 2025, is listed as an active Y Combinator Summer 2026 company with a two-person team, and publicly launched the Photon-1 research result in July 2026; PitchBook reports a $500,000 accelerator financing. There is no disclosed customer base or public product access, and the model is best described as pre-commercial in availability rather than definitively pre-revenue. Revenue data is conflicting and low-confidence: PitchBook leaves its current-revenue field blank, while Latka estimates approximately $220,000 in annual revenue, so commercial traction cannot be established from public evidence.
Founders & Leadership
Funding History
Y Combinator
Recent News
Induction Labs introduced “imagination models,” a foundation-model architecture that learns to predict future states from internet video. Its first model, Photon-1, is described as a sparse 106B-A5B MoE transformer trained on 18 years of computer-demonstration video; the company reports that it outperforms a production LLM on internal computer-use benchmarks while using 30× less training compute and costing 3× less to serve.
Y Combinator profiled Induction Labs as a San Francisco AI company founded by Jonathan Li and David Li. The profile describes the company’s research launch around imagination models and Photon-1, which learns from internet video and demonstrated computer use.
Induction Labs announced Axiom 1, a unified computer-use foundation model designed to see software, plan, and act directly. The company reports state-of-the-art OSWorld-Verified performance, surpassing listed GPT-5, o3, and Claude 4 Sonnet results, with 16× lower time per action than GPT-5.
Active Roles
0No active roles right now.
Get notified when they postBusiness Model
No public pricing, subscription, licensing, or customer-revenue model was identified. The latest materials present Induction Labs as a research lab seeking collaborators and talent while scaling its models, so its commercial revenue streams remain undisclosed.