Mahesh Sathiamoorthy, left DeepMind to build the rooms AI agents practice in. Mountain View, California.
A small applied-AI lab building the datasets, benchmarks, and training environments that frontier labs quietly rely on - and a $40M bet that reliability, not scale, is the next frontier.
Bespoke Labs is a research lab, not a chatbot company. It builds the underlying material other AI systems learn from: synthetic datasets scrubbed for factual grounding, benchmarks that measure whether an agent can actually finish a task, and - increasingly - simulated environments where an AI agent can fail safely before it fails in production. The company's argument is straightforward: agents break down not because models lack intelligence, but because they've never practiced in anything resembling the real world.
"Agents are unreliable. That single fact limits how long they can operate autonomously."
It's an unglamorous thesis for an unglamorous layer of the AI stack. But it is precisely the layer that Anthropic, OpenAI, Google DeepMind, Meta, and Amazon have quietly been drawing from.
A compact model built for one job: telling you whether a claim is actually grounded in its source document. It tops the public LLM-AggreFact leaderboard, runs in roughly 200 milliseconds, and beats GPT-4o and Claude 3.5 Sonnet at the task despite its size.
An open reasoning dataset built with the DataComp research community. OpenThoughts-114k and its successor, OpenThoughts3-1.2M, each became Hugging Face's #1 trending dataset and have been downloaded more than 500,000 times.
An open-source Python library for generating and curating synthetic training data, with built-in retrieval-augmented fine-tuning, batch inference, and data-quality scoring.
A benchmark for agentic coding and terminal use, used by Anthropic, OpenAI, and Google DeepMind to evaluate how well frontier models handle real, multi-step tasks.
An evolutionary prompt and policy optimizer that reportedly reaches strong results with a fraction of the rollouts required by standard reinforcement-learning methods like GRPO.
The company's newest line of work: composing realistic simulated worlds - codebases, microservices, sandboxed execution - where agents can be trained and tested at scale.
Bespoke Labs was founded in 2024 by Mahesh Sathiamoorthy, a former Staff Research Engineer at Google DeepMind who worked on large language models and recommender systems, and Alex Dimakis, a UC Berkeley professor and researcher in generative AI. The pairing reflects the company's split identity: one foot in shipped industry infrastructure, one foot in open academic research.
That academic instinct shows up in the company's product strategy. Instead of locking its best work behind a paywall, Bespoke Labs has repeatedly published its datasets and benchmarks for free, then built commercial services around the infrastructure needed to use them at scale.
Founded: 2024
HQ: 800 W El Camino Real, Mountain View, CA
Team: ~42 employees
Total funding: $40M
Latest round: Series A, $31.75M, July 2026
Lead investor: Wing Venture Capital
Industry: AI / Information Technology & Services
| Round | Amount | Date | Key Investors |
|---|---|---|---|
| Seed | $8.25M | 2024 | 8VC, with Jeff Dean, Spiros Xanthos, Dheeraj Pandey participating |
| Series A | $31.75M | July 2026 | Wing Venture Capital (lead), Mayfield, The House Fund, plus angels from Anthropic, OpenAI and Meta |
Stated use of funds: expanding the research team, scaling environment-building infrastructure, and growing commercial momentum with enterprise customers.
Bespoke Labs doesn't compete for consumer attention. It competes for a spot in the training pipeline of companies that do. That puts it alongside synthetic-data and data-curation vendors like Scale AI, Snorkel AI, Gretel, Together AI, Galileo, and Patronus AI - and, in the newer reinforcement-learning-environment space, alongside younger entrants like Mechanize and AfterQuery.
Open-source datasets and benchmarks first, commercial infrastructure second. Small, specialized models built for one job rather than general-purpose scale. A thesis built around environments and reliability, not raw model size.
Most competitors sell proprietary data-labeling or curation pipelines directly to enterprises, with less emphasis on public releases or open benchmarks that the research community can build on for free.
Mahesh Sathiamoorthy and Alex Dimakis found the company in Mountain View, closing an $8.25M seed round led by 8VC.
The company's first model release tops the public Grounded Factuality leaderboard.
Released with the DataComp research community; becomes Hugging Face's top trending dataset.
Anthropic, OpenAI, and Google DeepMind begin citing Bespoke Labs' benchmark and optimizer tools.
A larger successor dataset repeats the feat, cementing the project's reach in the open research community.
Bespoke Labs raises a combined $40M, repositioning around reinforcement-learning environments for reliable agents.
"The datasets aren't the product. They're the proof that we know what we're doing."
It builds tools, datasets, and training environments that make AI models and agents more reliable - spanning synthetic data curation, hallucination detection, and reinforcement-learning environments for agentic AI.
CEO Mahesh Sathiamoorthy, a former Google DeepMind research engineer, and Chief Scientist Alex Dimakis, a UC Berkeley professor.
A total of $40M: an $8.25M seed round led by 8VC and a $31.75M Series A led by Wing Venture Capital, announced in July 2026.
A compact model built to detect hallucinations by checking whether a claim is grounded in its source document. It tops the public LLM-AggreFact leaderboard and outperforms larger models like GPT-4o.
Its open datasets and benchmarks are reportedly used by Meta, Amazon, the Allen Institute for AI, Anthropic, OpenAI, and Google DeepMind, alongside a stated base of 200+ research teams and enterprise customers.
Reporting based on public sources including company statements, press releases, and third-party coverage as of July 2026. Figures such as team size and revenue are approximate where noted.