Proximal raised a $15M seed round at a $300M valuation led by General Catalyst. SV Angel, Diede van Lamoen of Go Global Ventures, and Chemistry participated, along with angel investors Liam Fedus, Kevin Weil, and Erik Bernhardsson.

The round closed in late 2025 and was announced in September 2026.

Calvin Chen and Justus Mattern co-founded Proximal in the fall of 2025. Mattern previously co-founded Revideo, an open-source animation library, and worked on open RL infrastructure at Prime Intellect.

The team includes employees from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel, and Jane Street. Former founders account for more than half of Proximal's headcount.

A frontier lab or enterprise customer works with Proximal's team to identify coding capability gaps in their models. Proximal's systems then turn source code, agent traces, and workflow artifacts into evaluations and targeted post-training data used to close those gaps.

Proximal runs agentic rollouts across execution environments ranging from single containers to Kubernetes clusters and multi-GPU setups, measuring task difficulty and performing quality assurance on generated data. This infrastructure lets Proximal test whether synthetic tasks are hard enough to push the capability frontier and whether model outputs meet quality standards.

Proximal also researches fuzzy verifiers, agents that assign rewards for code quality and maintainability rather than only functional correctness, to improve RL training signals.

Proximal publishes FrontierSWE, an open coding benchmark for ultra-long-horizon problems allowing up to 20 hours of agent execution per task. Tasks cover compiler optimization, ML research, and full system reimplementation, testing the types of problems targeted by Proximal's data products. Proximal developed FrontierSWE in collaboration with Modular, Prime Intellect, and Thoughtful Lab.

Proximal is researching ways to index public code, generate realistic codebases with millions of lines, and train multi-agent systems on tasks that take days or weeks to complete. These projects aim to provide a continuing supply of difficult training problems as existing tasks become easier for new models.

Proximal sells coding training data and post-training infrastructure to AI labs and enterprises. Contract terms and pricing units have not been disclosed.

Proximal's approach automates the repeated manual work of writing training tasks and crafting evaluations, shifting costs toward engineers, compute, and verification infrastructure rather than large teams of human labelers. Actual unit costs and margins remain undisclosed, and automation does not wholesale replace human expertise; engineering judgment is still required to identify capability gaps and validate data quality.

Proximal uses its own RL infrastructure to train models on its data products. This cycle of building data, training models, evaluating results, and iterating allows Proximal to validate data quality before delivery.

Proximal publishes open research such as FrontierSWE, which demonstrates its understanding of the capability frontier. Open benchmarks give potential buyers visibility into where current models fall short.

Proximal competes most directly with Scale AI and Surge AI, which offer coding datasets and RL environments for training frontier models. All three serve AI labs building coding agents, but they differ in data production methods and degree of automation.

Scale AI introduced RL Environments in early 2026, pairing simulated applications for training and evaluating agent behavior across tool use, computer use, and coding with expert-designed training data and verifiers. Scale combines machine learning pre-labeling with human review from its contractor workforce and has relationships with frontier labs across products spanning automotive labeling, GenAI data engines, and enterprise AI platforms.

Surge AI sells off-the-shelf frontier data, evaluations, and RL environments built by expert contributors who also work on custom engagements for frontier labs. Surge reported a 7.1 percentage point gain on SWE-Bench Pro for GLM-4.7 after training on its coding dataset. Under Surge's trusted-program model, partners can train first and pay if the data works.

The central question in this market is whether automated research pipelines can produce training data that is both harder and more useful than expert-built data. Scale and Surge both use human experts augmented by automation, while Proximal uses automated systems augmented by a small research team. As models improve, fewer humans can meaningfully judge or improve frontier model outputs, which could favor Proximal's approach, but buyers still need to assess whether the resulting tasks produce improvements in their own models.

Proximal has identified scientific domains such as drug discovery as potential expansion markets, where its core loop of converting unstructured data into evaluations, generating targeted post-training data, and improving model deployment could apply to molecular and biological problems. Entering these markets would require new data sources, domain-specific verifiers, and different execution environments, with Proximal's synthetic data generation and RL training infrastructure serving as a starting point.

Proximal has identified specialized chip design as another expansion market, where AI could assist with workloads that currently require teams of expert engineers. Hardware description languages and chip verification workflows resemble software engineering in their large codebases, deterministic test suites, and measurable performance metrics. Entry would require adapting Proximal's code-focused infrastructure to hardware-specific toolchains and simulation environments.

Proximal has identified legacy software used by hospitals, power grids, and government systems as a target domain. Modernization requires understanding large, poorly documented codebases and rewriting them safely, similar to the ultra-long-horizon coding tasks Proximal benchmarks through FrontierSWE. Unlike the scientific and chip design opportunities, this remains within software engineering but extends Proximal's customer base from AI labs to enterprises and governments operating critical legacy systems.

Training demand concentration: The pool of buyers for frontier coding training data is inherently small, limited to well-funded AI labs and large enterprises. Proximal has not disclosed customer concentration, but a pullback in training budgets or a shift toward insourcing data pipelines by a major buyer could materially reduce demand.

Reward hacking: Automated data and evaluation pipelines risk producing tasks that models learn to game rather than genuinely solve, undermining the usefulness of Proximal's training data. Detecting and correcting these failure modes requires ongoing research investment with uncertain returns.

Expansion economics: Moving from coding into scientific workflows, chip design, and legacy software requires different data sources, verifiers, and domain expertise. Each vertical carries its own acquisition costs and infrastructure demands, and pursuing them simultaneously could dilute the engineering focus behind Proximal's early traction.