What does Induction Labs actually build?

Induction Labs is a 2-person San Francisco research lab building foundation models that learn computer-use from unlabeled screen-recording video. Its flagship model, Photon-1, is a 106B-parameter, 5B-active MoE transformer, unveiled via YC Launch on July 23, 2026.

Is Induction Labs profitable or generating revenue?

No. As of the July 27, 2026 evidence compilation, Induction Labs is pre-revenue with no disclosed customers, pricing, or product API. Its only public evidence is a self-reported internal benchmark comparison from its July 23, 2026 launch post.

Could OpenAI or Google replace Induction Labs?

Yes, structurally. OpenAI's Computer-Using Agent, Google's Gemini 3.5 Flash native computer-use (announced May 19, 2026), Anthropic's Claude Computer Use beta, and Microsoft Copilot Studio (GA May 13, 2026) already ship the same capability as native features, with far more capital and distribution.

What is Induction Labs' biggest competitive threat?

Frontier labs folding computer-use into flagship products for free, plus $1B-class world-model rivals like Yann LeCun's AMI Labs (raised ~$1.03B, March 2026) and Fei-Fei Li's World Labs (~$1B), which compete for the same talent, compute, and investor attention.

Is Induction Labs' efficiency claim verified?

No. The company claims Photon-1 uses about 30x less pretraining compute and delivers 3x cheaper inference than Google's Gemini 3.1 Flash-Lite, but this is a vendor-claimed, self-reported internal benchmark against a low-tier model, not independently verified as of July 23, 2026.

How big is Induction Labs' team?

YC's directory lists Induction Labs' team size as 2 as of its July 2026 listing, though the company was actively recruiting engineers and researchers per its July 23, 2026 launch post; a third-party hub lists a broader 1-10 employee range.

YC Summer 2026 · United States

Induction Labs

2-person YC lab building a foundation model that learns computer-use from screen-recording video.

Little protection left

Assessed 27 July 2026 · Evidence as of 27 July 2026

Not much stands between this business and its pressures.

On the recordSince 27 July 2026 · Checkable from H2 2026 · Every call on the record Checkpoint: Independent benchmark or funding disclosure that tests the efficiency claim against flagship models

The call

Induction Labs is a tiny research team claiming an efficiency edge inside a category that OpenAI, Google, Anthropic, and Microsoft have already shipped as a standard feature.

Induction Labs is trying to build a general-purpose AI model that can control computers and automate tasks. The problem is that OpenAI, Google, Anthropic, and Microsoft have all already released computer-control features built into their own AI models as of mid-2026, and there is nothing stopping them from making their versions better. The one thing that could matter is if Induction Labs can prove their model truly runs 30 times faster and cheaper than the giants' models and convince customers to switch, but so far that speed claim is unverified and hard to believe given how quickly the big companies copy each other's techniques.

Why

AI pressure: OpenAI Operator/CUA, Google Gemini Computer Use, Anthropic Claude Computer Use, Microsoft Copilot Studio. Every frontier lab already ships computer-use natively, so Induction is a small model competing with the giants' side feature. Could AI do the core job? Yes, likely.

Bigger players: OpenAI, Google DeepMind, Anthropic, Microsoft. These labs have compute, distribution, and GA products in the same category Induction just launched into. Could a giant copy it? Yes, easily.

What protects it. An unverified claim of 30x cheaper pretraining and 3x cheaper inference than Gemini Flash-Lite, if it holds up.

What would change this call

Independent benchmark tests Photon-1 against flagship models like Gemini 3.1 Pro, not just Flash-Lite. Would show whether the claimed 30x efficiency edge holds against real competitors, not a weak baseline.
Induction Labs announces a funding round, product launch, or paying customer within 12 months. Would show the company is converting research into a viable business rather than staying pre-revenue.
OpenAI, Google, Anthropic, or Microsoft ship an efficiency or capability update to their computer-use features. Would show whether frontier labs are closing any gap Induction Labs claims to have.

Key numbers

Team size: 2 people per YC listing (2026-07)
Photon-1: 106B params, 5B active, MoE (2026-07-23)
Claimed 30x less pretraining compute vs Gemini Flash-Lite (2026-07-23)

Named risks

Frontier labs already ship native computer-use as free features, eroding Induction Labs' core differentiation before revenue. Severity: high
Sole claimed asset, a compute-efficiency edge, is self-reported and unverified against flagship (not Flash-Lite) models. Severity: high
$1B-class rivals (AMI Labs, World Labs) absorb talent, compute, and investor attention in the same thesis space. Severity: high
Unresolved YC batch discrepancy (S25 vs Summer 2026) and no disclosed funding amount raise transparency questions. Severity: medium
2-person team must scale hiring and compute rapidly to compete, straining execution capacity. Severity: medium

Rivals, as read in July 2026

Rival	Direction	Evidence
OpenAI (Computer-Using Agent/Operator)	losing to	Native computer-use agent shipped as production feature prior to Induction Labs' July 2026 launch
Google DeepMind (Gemini)	losing to	Gemini 3.5 Flash gained native computer-use at I/O, May 19, 2026, folding the category into flagship model for free
Anthropic (Claude Computer Use)	losing to	Public beta computer-use capability active as of May 2026
Microsoft (Copilot Studio)	losing to	Reached general availability with computer-use, May 13, 2026
AMI Labs (Yann LeCun)	losing to	Raised ~$1.03B at ~$3.5B pre-money valuation in same learn-from-observation thesis, March 2026
World Labs (Fei-Fei Li)	losing to	Reported ~$1B raised in same world-model/video-learning thesis space, March 2026

Questions people ask

What does Induction Labs actually build?: Induction Labs is a 2-person San Francisco research lab building foundation models that learn computer-use from unlabeled screen-recording video. Its flagship model, Photon-1, is a 106B-parameter, 5B-active MoE transformer, unveiled via YC Launch on July 23, 2026.
Is Induction Labs profitable or generating revenue?: No. As of the July 27, 2026 evidence compilation, Induction Labs is pre-revenue with no disclosed customers, pricing, or product API. Its only public evidence is a self-reported internal benchmark comparison from its July 23, 2026 launch post.
Could OpenAI or Google replace Induction Labs?: Yes, structurally. OpenAI's Computer-Using Agent, Google's Gemini 3.5 Flash native computer-use (announced May 19, 2026), Anthropic's Claude Computer Use beta, and Microsoft Copilot Studio (GA May 13, 2026) already ship the same capability as native features, with far more capital and distribution.
What is Induction Labs' biggest competitive threat?: Frontier labs folding computer-use into flagship products for free, plus $1B-class world-model rivals like Yann LeCun's AMI Labs (raised ~$1.03B, March 2026) and Fei-Fei Li's World Labs (~$1B), which compete for the same talent, compute, and investor attention.
Is Induction Labs' efficiency claim verified?: No. The company claims Photon-1 uses about 30x less pretraining compute and delivers 3x cheaper inference than Google's Gemini 3.1 Flash-Lite, but this is a vendor-claimed, self-reported internal benchmark against a low-tier model, not independently verified as of July 23, 2026.
How big is Induction Labs' team?: YC's directory lists Induction Labs' team size as 2 as of its July 2026 listing, though the company was actively recruiting engineers and researchers per its July 23, 2026 launch post; a third-party hub lists a broader 1-10 employee range.

Timeline

2026-07-24Digg covers Induction Labs' Photon-1 imagination models for video learning
2026-07-23Induction Labs publicly unveils Photon-1 via YC Launch post
2026-05-19Google announces Gemini 3.5 Flash native computer-use at I/O
2026-05-13Microsoft Copilot Studio reaches general availability with computer-use
2026-05Anthropic Claude Computer Use reaches public beta
2026-03-09Yann LeCun's AMI Labs raises ~$1.03B for world-model thesis
2026-03Fei-Fei Li's World Labs reported to have raised ~$1B
2025-10Google Gemini 2.5 Computer Use enters developer preview
2025-07-15Induction Labs incorporated as California Stock Corporation

Revision history

27 July 2026

First assessed. Little protection left. The text above is fixed as of this date and is not edited afterwards.

In the same band

Cerenovus Belvedir Hoplite