Understudy Labs wants you to stop renting AI - and own the model instead.

The pitch fits on a bumper sticker: Don't Use Their Models. Use Yours. Behind that line is a bet about where artificial intelligence is heading - not toward ever-bigger models rented by the token, but toward smaller, task-shaped models that companies actually own. Understudy Labs, a two-person startup in San Francisco and a member of Y Combinator's Summer 2026 batch, is building the machinery to make that switch feel safe.

The name is a theater metaphor, and it holds up. A frontier model - Claude, GPT, Gemini - plays the lead. It's brilliant, expensive, and booked solid. Understudy trains a replacement in the wings: a smaller open-weight model that studies the star's performance, learns the part, and goes on only when it can prove it won't drop a line. The lead never knows it's being replaced. The audience never notices. The bill shrinks.

01 / The ProblemThe cost wall is real

Every company building with large language models eventually meets the same unpleasant arithmetic. The demo is magic. The monthly invoice is not. As usage grows, so does the dependency - on someone else's model, someone else's pricing, someone else's roadmap. Understudy frames its purpose plainly: bring relief to AI-native companies hitting the cost wall, and bring ownership to enterprises moving from experiments into production.

The insight underneath is one both founders learned at scale: a general-purpose model is overkill for most production jobs. If your agent classifies support tickets, extracts fields from invoices, or scores sentiment all day, you are paying frontier prices for a fraction of frontier capability. A specialist that does that one job at 95 percent of the quality and a twentieth of the cost usually wins.

02 / How It WorksCapture, evaluate, train, deploy

Understudy's product is a loop, not a one-time migration. It starts open source: a toolkit that installs into the coding agents teams already use, with hosted infrastructure kept optional so data can stay in-house.

The catch that makes it interesting: production data feeds back into training, so the model you own compounds over time.

That last step is the quiet centerpiece. Most cost-cutting in AI is a leap of faith - swap the model, cross your fingers, hope quality holds. Understudy refuses to deploy anything that loses the eval. It sells confidence as much as savings.

03 / The NumbersWhat the successor can do

The company is in private preview with design partners, so its figures should be read as early, task-specific results rather than universal guarantees. They report a fine-tuned understudy scoring 13 percent higher on eval than Claude Sonnet 3.5 on a target task, matching roughly 1.13x its performance at about a quarter of the cost.

04 / The PeopleAn ad-ranking pedigree

The playbook comes straight from ad-tech, where models serving billions of requests per day live or die on cost per prediction. Luis Manrique, co-founder and CEO, joined Google in 2013 through the Wildfire acquisition, spent years in ads ML, was a principal PM on Instacart's Ads and Carrot AI, and was an early operator at Gumloop through its growth. He holds ten patents in ML and AI-agent infrastructure.

Aamir Poonawalla, co-founder and CTO, spent about ten years at Instacart leading ads serving and infrastructure, and helped scale its Curbside Pickup business from zero to roughly four billion dollars in gross transaction value. He is a repeat Y Combinator founder (W2012) with a master's in computer science from Georgia Tech. The pair are backed by Instacart co-founder Max Mullen and Gumloop co-founder Max Brodeur-Urbas.

05 / The MarketWhere Understudy fits

The AI infrastructure aisle is crowded. Managed fine-tuning and distillation platforms like OpenPipe, Predibase, Together AI, and Fireworks will also shrink your model. Evaluation and observability tools like LangSmith and Braintrust will help you measure it. And the simplest competitor of all is inertia - staying on the frontier API you already pay.

Understudy's two-part wedge is ownership plus a gate. The output is an open-weight model you keep and run on your own infrastructure, and the switch is blocked unless it wins an evaluation you defined. For teams that have been burned by silent quality regressions, that second part is the pitch. Who is it for? Data-rich teams with real task volume, cost or latency pressure, and domain experts who can say what "good" looks like.

Watch & demo: Understudy Labs had not published a public product demo or founder interview video at the time of writing. Check the website and LinkedIn for the first release.