Labelbox started in 2018 by labeling data. You give a model enough good examples, the thinking went, and it gets smarter. The San Francisco company’s new raise suggests the job has changed. On Oct. 14, Labelbox said it had raised more than $100 million in Series E financing, a mix of equity and a credit facility, at a $1.25 billion pre-money valuation.
Infinitum Partners led the round. Wipro Ventures, Andreessen Horowitz, B Capital, Tau Ventures and Snowpoint Ventures also took part. Chief executive and co-founder Manu Sharma said the money will fund research and engineering for two core products, expand the customer teams, and secure access to compute for the next phase. Compute, in his telling, sits at the center of what comes next.
Practice, not just examples
The first product is Horizon. It builds reinforcement learning environments and evaluation for frontier labs. Picture a simulated company, filled with code, dashboards, spreadsheets and documents. A model steps inside and tries a multi-step task. It might chase a figure across files, check it against a dashboard, then draft the next move. Every attempt gets recorded. Every attempt gets graded.
That grading is the point. A single right or wrong answer teaches a model very little about a long job. A scored run shows where it drifted, which step it skipped, and what a better path looked like. Do that thousands of times and the model is not memorizing. It is practicing.
That shift matters beyond the labs. Readers who followed how Google opened its best AI models to 200 European researchers will recognize the same pattern. Access is only step one. What researchers and companies do with repeated, measured trials is where the gains compound.
Agents that work inside real software
The second product is Recursion Managed Agents. Labelbox opened it to enterprise customers earlier this month, in October 2026. Instead of asking a company to rip out its tools, the agents deploy inside existing business applications. They are scored against rubrics, clear written standards for what a good result looks like, and they improve with each run, according to the company.
It is a modest claim with a big implication. Most office work already lives in familiar software. An agent that learns inside those screens, judged against a rubric a manager can actually read, is easier to trust than a black box that arrives with a promise. You can see the score. You can argue with the rubric. You can fix both.
Why rubrics change the conversation
A rubric turns a vague complaint – the output felt off – into a specific one. The summary missed two line items. The tone was wrong for a client note. The agent skipped the approval step. Once the failure has a name, the next run can target it. That is how human trainees improve, and Labelbox is betting it is how agents will too.
The same discipline shows up in other corners of the field. Coverage of Gemini 4 Argon arriving for hard problems pointed to models being judged on tougher, longer tasks. Horizon is built for exactly that kind of judging, in settings that look like work rather than a quiz.
A different kind of funding story
Labelbox reports more than 400 customers, including over 90 percent of leading US AI labs. It also reports profitability in 2026 and growth of more than 5x year over year. Those figures come from the company’s own announcement and founder note, so treat them as company claims, not audited results. Even with that caution, the shape of the story stands out.
Profitability alongside fast growth is not the usual script for a company raising nine figures. Plenty of firms raise because they have to. Labelbox, on its own numbers, raised from a position of strength, and it is spending the money on depth – research, engineering, customer teams and compute – rather than on a splashy land grab.
There is a civic version of this idea too. The recent $50 million Cities Deliver push to put AI to work for residents rests on the same logic. Tools earn trust when people can watch them work, score the result, and feed the lesson back in.
What to take away
For lab teams, the practical lesson is simple. Budget for practice, not just for data. If your model will handle multi-step work, it needs environments where that work can fail safely, get graded, and run again. A folder of perfect examples will not teach recovery.
For enterprise buyers, ask two questions before you sign anything. What rubric will the agent be scored against, and who wrote it? And can you see the scores improve run by run, inside the applications your staff already use? If a vendor cannot answer both, you are buying a demo.
For everyone else, watch the language. The field spent years talking about training data as if it were fuel – pour in more, go further. Labelbox is talking about something closer to coaching. Fuel still matters. But a $1.25 billion valuation, and more than $100 million of fresh capital, now sits behind a plainer idea. Models get good the way people do. They try, they get marked, and they try again.





































