Primus
Give it a research question and get back a coded experiment, run results, and a finished paper, roughly 30x faster than doing it yourself.
The verdict on Primus: ML researchers and R&D teams who want a full hypothesis-to-paper research loop automated What separates Primus from most autonomous research demos is that it shipped receipts: 30 completed papers in 30 days, with at least one citation from a major lab, not just a benchmark screenshot. Pricing: Free during waitlisted rollout, GPU-limited.
Best For
ML researchers and R&D teams who want a full hypothesis-to-paper research loop automated
Standout Feature
Output has already been cited in published research from Google DeepMind, not just a demo claim
TL;DR
The most credible autonomous research tool available right now, if you can get off the waitlist.
Alternatives
Fit Score: use-case reference
How this is scoredPrimus scored 1 to 5 on five dimensions for the named use cases below. Part of our catalog-wide use-case reference, not a product-quality review rating.
ML researchers and R&D teams who want a full hypothesis-to-paper research loop automated
4.0/5- Capability
- 5/5
- Reliability
- 3/5
- Value
- 5/5
- Ease
- 3/5
- Compatibility
- 4/5
Output has already been cited in published research from Google DeepMind, not just a demo claim, and it's free to access during the current rollout with genuinely broad domain coverage (standout, pros). Reliability and ease are capped by a waitlist with limited GPU capacity and a real requirement for domain expertise to validate output (cons).
Non-researchers who need to evaluate or trust output without domain expertise
2.0/5- Capability
- 2/5
- Reliability
- 2/5
- Value
- 3/5
- Ease
- 1/5
- Compatibility
- 2/5
Its own cons state it requires domain expertise to validate output before real use, and someone without domain expertise will struggle to know if the output is actually sound (ourTake), making it a poor fit for non-researchers.
Editorial judgment grounded in this tool's own published pricing, features, and our existing pros and cons, not a certification or a measured benchmark. Product quality is assessed separately under our review methodology.
Overview
Not yet reviewedProduct quality is assessed from original evidence. How reviews work.
Primus, built by Transformer Lab, is an autonomous AI researcher that runs the entire scientific research loop on its own: it forms a hypothesis, reads millions of relevant papers, writes and runs the experiment code on real compute, iterates on the results, and writes up a finished paper, roughly 30 times faster than a human research team working the same problem. It went live August 10, 2026 with gradual onboarding from a waitlist due to limited GPU capacity, and it's free to use during this rollout. In a 30-day internal test, Primus produced more than 30 complete papers on previously unanswered questions spanning machine learning, seismology, physics, crystallography, and biology, and findings it helped produce have already been cited in published research from labs including Google DeepMind.
It isn't limited to a narrow problem type: you can ask it to fine-tune a model, write and benchmark a custom GPU kernel, explore open questions left in a paper you supply, or test a new forecasting approach against your own sensor data. The output is a real deliverable, code, benchmarks, and a written paper, though it's still built for people who can evaluate research output, not a black box for non-researchers.
The source of the legacy catalog rating and review count has not been substantiated. Neither is presented here as a verified review rating or included in review structured data.
Workflow Fit Scores are separate, use-case-specific editorial judgments, not user-review averages. Read the Fit Score methodology.
- • Output independently cited in published research
- • Genuinely broad domain coverage, not narrowly scoped
- • Free to access during the current rollout
- • Waitlisted with limited GPU capacity
- • Requires domain expertise to validate output before real use
- • Mostly proven on ML problems so far despite broader claims
Key Features
- End-to-end autonomous research loop
- Reads millions of papers to ground hypotheses
- Writes and runs real experiment code
- Domain-flexible (ML, physics, biology, and more)
- Produces code, benchmarks, and a written paper
- Self-improving with more experiments run
Trust & Data
Verified 2026-09
- Training on your data
- NoThe privacy policy states plainly it does not sell personal data or use confidential research datasets to train publicly available base models, but that commitment is scoped specifically to public base models; the Terms' broader license to use data for operating, maintaining, and improving the services leaves open whether internal or private model fine-tuning could fall inside that improvement language, so treat the no-training claim as qualified rather than absolute. The lab.cloud domain is confirmed correct for this product; the homepage itself is titled 'Primus, the world's first fully autonomous AI researcher.'
- Compliance
- Not published
- Retention
- Retained only as long as necessary to fulfill the purposes collected or to comply with legal obligations; deletable by request at any time.
- Export
- Partial export
Sources: lab.cloud/privacy, lab.cloud/terms
As published by the vendor. Verify independently before purchase decisions.
Our Take
That's a meaningfully higher bar of proof than most agentic tools clear at launch. The honest limitation is that this is built for people who can already evaluate research quality, an ML engineer or R&D lead gets real leverage here, someone without domain expertise will struggle to know if the output is actually sound. The waitlist and GPU constraints mean you can't just sign up and start today, but it's worth getting in line if research throughput is a real bottleneck for your team.