Perplexity has launched Perplexity Frontier Research, a new initiative to build an AI system that can do research on its own. The company announced it on X on November 16, 2026. It is backing the effort with more than 10,000 GPUs and says the work will be released openly.
The team plans to publish papers, models, data and benchmarks, and to test its results inside Perplexity's products. It is hiring researchers and engineers in San Francisco, Palo Alto, New York and London.
What Perplexity Frontier Research is
Perplexity describes the goal as an automated scientist. Its research page describes science as a loop. You choose a question, read what is known, suggest an answer, run an experiment and update what you believe. The company wants that loop to run inside software, for anyone.
In Perplexity's picture, many agents share the work. Some search widely across thousands of sources, some go deep on one thread, and others propose experiments or critique ideas. The page says the system should know what it does not know and remember what it learns.
The company also sets a higher bar than search. It wants systems that produce new, understandable scientific insight, which it contrasts with black-box predictions. To get there, Perplexity has split the work into six areas.
The six research areas
Each area covers one part of the automated scientist:
- Multi-agent orchestration: getting many models to work together, with the best quality for any budget
- Pre-training: efficient open-weight models, new architectures and synthetic data
- Continual learning: models that keep learning after release and remember past work
- Reinforcement learning: training with verifiable or rubric-based rewards in environments built from real usage
- Systems: inference, training and secure runtimes for long-running agents
- Frontier evaluations: benchmarks that measure discovery instead of recall
Orchestration ties the others together. One example on the page is advisor escalation, where a small on-device model asks a larger one for help only when it needs to. Perplexity says it wants the best-quality AI service at the lowest cost.
Pre-training is the most notable item on the list. Perplexity has so far built its products on other companies' models, including its Sonar models, which it fine-tuned from open-weight models. Now it says it will pre-train its own efficient, specialized open-weight models. The person in charge of this agenda joined in August.
Who leads Frontier Research
The research lead is Andrew Gordon Wilson, a professor at New York University. He announced on August 24 that he was joining Perplexity, promising "paradigm shifting work" in continual learning and agent collaboration. He keeps his NYU research group.
Wilson is known for work on probabilistic machine learning and on why neural networks generalize. One of his papers, Deep Learning is Not So Mysterious, is among the example works on the new research page. CEO Aravind Srinivas said in August that Wilson reports to co-founder Denis Yarats.
Yarats said then that Perplexity had been building infrastructure for large-scale reinforcement learning. He promised a research agenda and open-source releases. The November launch is that agenda, and some of the releases are already out.
What Perplexity has already released
The research page lists open work the team will build on. It includes:
- Decider v1.2, a 27-billion-parameter open-weight model that returns probabilities for decisions instead of text
- Contextual and multimodal embedding models for search, also with open weights
- SPACE, isolated runtimes that let agents pause and resume work without access to credentials or the host
- WANDR, a benchmark for research agents that must search both wide and deep
- pplx-garden, with GPU communication kernels and Lily, an inference engine for Apple silicon
One reply under the announcement noted that Decider and the embedding models are built on Alibaba's Qwen. That makes pre-training the clearest change from what Perplexity has shipped so far. Perplexity is also entering a field where larger labs already have a head start.
How Perplexity compares with other AI scientist projects
Perplexity is not the first to chase an automated scientist. Sakana AI released The AI Scientist in 2024, and in 2025 a paper written by its second version passed peer review at an ICLR workshop. Google introduced its AI co-scientist, a multi-agent system built on Gemini, in February 2025.
OpenAI has the most public timeline. In October 2025, Sam Altman set two goals. OpenAI aimed for an automated AI research intern by September 2026 and a full AI researcher by March 2028. OpenAI, Google and Anthropic train their own frontier models on far more than 10,000 GPUs.
Perplexity is betting on a different mix. It combines orchestration of many models, smaller open models and a search product used by millions. The page calls that product "the most challenging test of all" for the research. Readers have already started asking how the team will pass it.
How people reacted
Replies under the announcement focused on two things: pre-training and evaluation. Several users asked whether Perplexity would build its own base model or keep post-training open weights from others. Others welcomed the promise to release models and benchmarks openly.
Evaluation drew the sharpest questions. One user asked how the team would tell whether a run found something new or just wrote a confident report. Another wanted a test where the starting premise is wrong and the agent has to drop it halfway. Several pointed to the risk of agents citing each other in a loop.
Perplexity's page names the same problem. It says new evaluations must separate reasoning from memorization and real understanding from good predictions. Some users were more practical and asked when better sources or higher usage limits would reach the app.
Can Perplexity build an automated scientist
Not yet, and Perplexity does not claim to have one. Frontier Research is an agenda with a lead, a GPU budget, a hiring drive and a set of early open releases. The automated scientist is the stated destination.
The plan has real strengths. Perplexity runs a search product at scale, has a respected academic in charge and already publishes open models and benchmarks. Its focus on cheap orchestration also fits a company that does not train frontier models.
The gap is compute and proof. 10,000 GPUs is a serious budget, but it is small next to the clusters at OpenAI and Google. Perplexity will need to show results on evaluations that measure discovery. For now, Frontier Research is a credible start, and its first papers and models will show how far it can go.
Sources
- Perplexity Frontier Research announcement, Perplexity on X – November 16, 2026
- Perplexity Frontier Research, Perplexity – November 16, 2026
- Perplexity launches research initiative to build an automated scientist, Superpower Daily – November 16, 2026
- Perplexity hires NYU professor Andrew Gordon Wilson to lead research, RuntimeWire – August 24, 2026
- Perplexity Decider v1.2, Hugging Face