Samaya AI

Samaya AI is an American artificial intelligence company that builds AI agents for financial research and analysis, used by Morgan Stanley and a number of hedge funds.^[3] It was founded in 2022 by a group of AI researchers from Google Brain, Meta AI, Amazon Web Services and the Allen Institute for AI, and is led by co-founder and chief executive Maithra Raghu, a former Google Brain research scientist.^[3][8] Its customers include Morgan Stanley, whose Institutional Securities Group uses the platform.^[3][18] Samaya raised a $43.5 million Series A led by New Enterprise Associates in May 2025, and NVentures (NVIDIA's venture arm) and Databricks Ventures invested in February 2026.^[3][14][15]

Samaya also publishes research. Its staff co-authored the 2023 long-context paper "Lost in the Middle", and in 2026 the company released FrontierFinance, an open benchmark of 220 expert-written investment research questions.^[13][22] On 28 October 2026 it announced Samaya Sterling, a post-trained version of NVIDIA's Nemotron 3.5 Lightning (30 billion total parameters, 3 billion active). Samaya reports that Sterling scores 58.9% on FrontierFinance at $0.61 per question. On Samaya's leaderboard, the best frontier model, Claude Fable 5.1, scores 55.9% at $6.89.^[1][24]

Overview

History

Founding (2022)

Samaya says it was founded in 2022, "before ChatGPT", on the view that "expert intelligence emerges from specialization."^[7] Fortune and TIME both describe the founders as researchers from Google Brain, Meta, Amazon and the Allen Institute for Artificial Intelligence.^[3][8] Raghu told Fortune that she and her co-founders were watching generative AI develop inside those labs and saw clear uses in financial services, but believed that products specialized for one domain would beat general-purpose large language models.^[3] TIME names Fabio Petroni as Raghu's co-founder.^[8] Petroni worked as a research engineer and manager at FAIR, Meta AI in London from 2018 to 2022. His European Molecular Biology Laboratory (EMBL) profile lists him as a Samaya co-founder since 2022 and as an EMBL group leader since September 2025.^[12]

Early research (2023-2024)

Three Samaya researchers (Ashwin Paranjape, Michele Bevilacqua and Fabio Petroni) co-authored "Lost in the Middle: How Language Models Use Long Contexts" with Nelson F. Liu, Kevin Lin, John Hewitt and Percy Liang. The paper lists their affiliation as Samaya AI, alongside Stanford University and UC Berkeley.^[13] It was first posted to arXiv in July 2023 and accepted by Transactions of the Association for Computational Linguistics. It found that language models use information best when it sits at the beginning or end of a long input context and worse when it sits in the middle.^[13] The finding is often discussed alongside long-context retrieval tests such as Needle in a Haystack. Later Samaya papers covered instruction-following retrieval (Promptriever, 2024) and long-horizon agentic search ("Lost in the Maze", 2025).^[21][29][30]

Series A and Causal World Models (2025)

On 14 May 2025 Samaya announced $43.5 million in funding led by New Enterprise Associates (NEA). Fortune reported that the valuation was not disclosed, and FinTech Global noted that the announcement disclosed no earlier rounds.^[3][4][5] Fortune named Eric Schmidt, Yann LeCun, Two Sigma co-founder David Siegel and Sixth Street vice chair Marty Chavez as participants.^[3] NEA partner Tiffany Luck said Samaya fit NEA's thesis that AI tools must be built for specific industry verticals. NEA's announcement said Samaya was seeing 100% month-over-month growth in usage.^[3][4] Samaya later described the round as its Series A.^[7]

Samaya launched Causal World Models with the round. The agent had been in research preview. It builds a cause-and-effect graph of an economic question and reasons over that graph. In one example Samaya used it to model how the Trump administration's proposed tariffs would affect the U.S. economy.^[3][6] Fortune also reported Raghu's description of the company's first research tool as a "lattice of experts" architecture. In it, many smaller language models each handle part of a research task and some models check the output of others.^[3]

In August 2025, TIME named Raghu to its TIME100 AI 2025 list in the "Innovators" category.^[8][9] Samaya introduced a Deep Research Agent in July 2025 and announced data partnerships with Crunchbase (June 2025) and the expert network Third Bridge (January 2026).^[19][33]

Agent Control Plane and new investors (2026)

On 10 February 2026 Samaya launched the Agent Control Plane (ACP) and announced new investment from NVentures and Databricks Ventures. The amount was not disclosed.^[14][15] Samaya describes ACP as an architecture for building personalized, long-horizon financial agents that are configured through natural-language instructions and remain auditable. Its components include an integrated planner, long-horizon executor and memory module, a reasoning module, and context-management and optimized-tools modules.^[14] FinTech Global and Pulse 2.0 both reported the company's claim that it was in production with more than 10,000 professionals at one of the world's largest banks.^[15][16] Samaya's announcement also cited internal evaluations with enterprise clients in which generic LLMs reached 53% accuracy on financial reasoning tasks against 98% for Samaya. The announcement does not describe how that evaluation was run.^[14]

Later in 2026 Samaya added a PitchBook data integration (June), released FrontierFinance (July), and in September added filing and news alerts delivered through Model Context Protocol, a set of fifty prebuilt analyst workflows and a Daloopa integration.^[19][23] Samaya Sterling followed on 28 October 2026.^[1]

Leadership and team

Samaya's founding-team page lists four people as of 30 October 2026:^[11]

Raghu was a research scientist at Google Brain from 2015 to 2022, where she studied representations and learning in deep neural networks. She received her PhD in computer science (machine learning) from Cornell University, advised by Jon Kleinberg, and studied mathematics as an undergraduate at Trinity College, Cambridge.^[10][9] Her academic papers include "Do Vision Transformers See Like Convolutional Neural Networks?" (NeurIPS 2021) and "Transfusion: Understanding Transfer Learning with Applications to Medical Imaging" (NeurIPS 2019), and she co-wrote a survey of deep learning for scientific discovery with Eric Schmidt.^[10] TIME wrote that her Google Brain work focused on specialized AI tools for healthcare.^[8]

Zhang is the first author and Paranjape the last author of the FrontierFinance paper.^[22]

Funding and investors

Samaya's investor page also lists Reid Hoffman, Nat Friedman, Jim Breyer, Diane Greene, Christopher Manning, Jeff Dean, Mark Cuban, Ram Shriram and Jeff Hammerbacher, without dates or amounts.^[17]

Products

Samaya sells a research platform for buy-side and sell-side professionals. Users can ask questions, delegate multi-step research to agents, and set alerts on filings, news and other sources. Answers carry citations that open the underlying passage.^[19][20] The platform searches company filings, earnings transcripts, published research and a firm's own notes and data.^[19] Samaya says it uses its own custom models, data index and retrieval engines.^[23]

Data integrations announced by Samaya include Crunchbase (private-market data, June 2025), Third Bridge (expert interviews, January 2026), PitchBook (June 2026), Finnhub (U.S. corporate bonds) and Daloopa (as-reported financials, September 2026). Several, such as PitchBook and Daloopa, depend on the customer's own subscription or provider access.^[19][33]

Customers

The customer Samaya names publicly is Morgan Stanley. Katy Huberty, Morgan Stanley's global director of research, said the bank would partner with Samaya "across all divisions of our Institutional Securities Group" to draw insights from its research library and external sources.^[3][18] Fortune also reported deployments at "a number of hedge funds" in May 2025.^[3] In June 2025 Samaya said it had "thousands of users globally".^[7]

Competitors

Fortune listed the startups Model ML, V7 Labs and Rogo as competitors in AI financial research, along with in-house efforts at JPMorgan and Bloomberg.^[3]

Research

Samaya publishes papers and research blog posts on retrieval, agents and evaluation. Its research page lists 13 papers from 2023 to 2026, including:^[21]

In an 8 October 2026 blog post, Samaya reported that recent frontier models running in its finance harness predicted company revenue one week before earnings with lower error than analyst consensus. The test covered 456 Q2 earnings releases and compared the models against both the raw consensus and a bias-corrected consensus.^[28] This is company research and has not been peer reviewed.

FrontierFinance

FrontierFinance is an open benchmark for finance AI agents, created by Samaya's machine-learning team and released on 9 July 2026.^[23] The technical report, posted to arXiv on 12 August 2026 (arXiv:2608.11683), is by Yuhao Zhang, O. Ozan Koyluoglu, Thejas Venkatesh, Richard Diehl Martinez, Vishank Bhatia, Arash Alidoust and Ashwin Paranjape.^[22] The dataset is on Hugging Face under CC BY 4.0, behind a gate that asks users not to reshare it, to limit contamination. The grading code is on GitHub under the Apache 2.0 license.^[25][26]

Design

The benchmark has 220 queries and 11,543 rubric items, an average of 52.5 per query. Every rubric is supported by publicly available data and tagged with a source category.^[22][23] The queries cover six use cases across an investor's workflow:^[24]

In-house finance experts drafted the queries, wrote rubrics under strict guidelines that Samaya iterates on, audited and pruned them, and rebalanced the set by use case, capability and difficulty. That process produced more than 4,500 annotated queries, from which the 220 public ones were drawn. Samaya keeps the rest for internal research and future dataset releases.^[23] Samaya says SEC filings support only 39% of rubrics. The rest come from transcripts, investor presentations, market data and professional knowledge.^[23]

The headline metric is the rubric qualification rate: the share of rubric items an answer satisfies, macro-averaged across queries. Each rubric judgment is a majority vote of three judge models (an LLM-as-a-judge setup).^[23] To compare difficulty with older benchmarks such as FinanceBench, Vals AI's Finance Agent v2 and BigFinanceBench, Samaya pooled their examples and had an AI agent judge pairs by difficulty. It then fit Bradley-Terry scores, and by that measure FrontierFinance is harder on average.^[23][24]

Systems are run in three kinds of harness: a model with its provider's built-in web search, the open-source Finance Agent v2 harness (six tools including the SEC EDGAR API, market prices, web search and a calculator), and Samaya's proprietary harness. Every harness connects only to public data.^[23] Results depend heavily on the harness. The paper finds that "the tool harness, not the model alone, strongly shapes quality and efficiency".^[22]

Results

At release, Samaya reported its in-house system at 50.8% and Claude Fable 5 at 49.2% as the best frontier model. The August technical report gave the in-house system 56.0%, and Kimi K3 at 46.4% as the best open-weight model.^[23][22] Screening and Discovery and Sector, Industry and Macro were the hardest use cases. The paper reports that even the best systems reached only 33% and 39% on them.^[22]

The top of Samaya's leaderboard as of 28 October 2026:^[24]

The leaderboard notes say Sterling's cost covers the research model, tools and helper models. Other systems' costs count only the agentic model and exclude search API and data-index charges.^[24] Samaya adapted the Finance Agent v2 harness with a 200-tool-call limit and a 300-second tool-call time limit, and ran every model with thinking on at its default reasoning level.^[24]

At least one other model developer has reported results on the benchmark. StepFun listed FrontierFinance among its finance evaluations for Step 5 Preview in September 2026. It reported 66.4% for Step 5 Preview, 69.7% for Claude Opus 5 and 62.6% for Kimi K3.^[27] Those numbers come from StepFun's own runs, not Samaya's leaderboard, where Kimi K3 scores 46.4% in the Finance Agent v2 harness. The two sets are not directly comparable.^[24][27]

Samaya Sterling

Samaya Sterling is a financial research agent model that Samaya announced on 28 October 2026. Samaya describes it as "a post-trained model for state-of-the-art financial research".^[1] The research post is credited to O. Ozan Koyluoglu, Richard Diehl Martinez, Thejas Venkatesh, Yuhao Zhang and Ashwin Paranjape.^[1] Raghu wrote that Sterling "started off as a two person research effort to see if we could increase efficiency in agentic orchestration".^[2] On the benchmark site, Samaya calls Sterling its new system and its earlier agent the "Legacy Samaya System".^[24] All Sterling figures below are Samaya's own measurements on its own benchmark and have not been independently reproduced.

Base model and harness

Sterling starts from NVIDIA's Nemotron 3.5 Lightning, an open-weights 30B-parameter mixture-of-experts model with about 3 billion parameters active per token, with reasoning ("thinking") enabled.^[1] Samaya trained it with LoRA adapters on Tinker, the fine-tuning API from Thinking Machines Lab.^[1]

The model runs inside an adapted version of Samaya's production harness, which has dozens of finance-specific tools, including data tools backed by in-house retrieval models.^[1] To fit a small model, Samaya removed detailed guidance from the system prompt and tool descriptions on the grounds that the model could learn it in training, and merged tools to shorten the prompt. The trained agent calls 37 tools.^[1] Two helper models manage context: a condenser shortens long tool results, and a compactor summarizes the trace as it grows. A research budget caps how long the agent works, after which a separate report writer produces the cited final answer from the evidence gathered.^[1] Samaya says the model and harness were "co-adapted" through training. Small harness changes moved scores in either direction, so it trains and evaluates in the harness it serves.^[1][2] The condenser, compactor and research budget are forms of context engineering, and Samaya tuned them together with the model during post-training.

Training questions came from an internal dataset written and graded by finance professionals using the FrontierFinance process. Samaya says FrontierFinance was strictly held out: no benchmark question was used in training, and checkpoints were selected on a separate pool of 100 internal validation questions.^[1]

Training

Training ran in four stages, two of supervised fine-tuning (SFT) and two of reinforcement learning with GRPO (group relative policy optimization):^[1]

In stage 2, the post reports that the best of the eight strategy answers met 80% of essential rubric items on average, against about 66% for the single best strategy.^[1] Stage 3 used the existing teacher traces as GRPO groups, so the model learned from the contrast between strong and weak strategies without generating new samples. The post compares this to Direct Preference Optimization and notes that it makes each step cheap.^[1] Stage 4 trained on questions where teacher scores varied. In the last four steps it also used questions where the agent most often ran out of research budget. Samaya applied truncated importance sampling to correct for the mismatch between sampling and training models, which it calls a known issue for mixture-of-experts models.^[1]

The two imitation stages added 12.3 percentage points and the two RL stages 4.1 more. The median number of tool calls per question rose from 3 for the base model to 17.^[1] On the hardest third of FrontierFinance, the score went from 31.8% to 45.2%, a 13.4-point gain. Samaya reports that hard questions improved most in the stages whose training pools contained more of them: synthetic SFT (384 of 968 questions hard) and online GRPO (30% hard after sampling).^[1]

Reported results

Source: Samaya.^[1][24]

Samaya's headline claim is that Sterling scores 3.0 points above Claude Fable 5.1 at roughly 11 times lower cost per question ($0.61 against $6.89).^[1][2] The comparison crosses harnesses: Sterling runs in Samaya's proprietary harness with its in-house retrieval, while Fable 5.1 was measured in the open-source Finance Agent v2 harness.^[24] Samaya says the research model itself accounts for $0.04 of Sterling's $0.61 per question, with the rest going to tool calls and the helper models. Frontier closed models doing the same research in the same Samaya harness cost $0.56 to $0.95 per question, but the post does not give their scores in that harness.^[1]

The post also reports a reward-shaping experiment. Samaya added a term to the online GRPO reward based on research-context length, zero up to 80,000 tokens and growing to 0.3 at 240,000 tokens. Used as a penalty, at step 20 it cut cost by 30% for 1.1 points less quality. Used as a bonus, it raised quality by 0.9 points for 115% more cost. Samaya presents this as a training-time control over the quality and cost trade-off.^[1]

Planned work

Samaya says it is exploring swarms of small agents running diverse research strategies, firm-specific models that encode a client's data, tools and past decisions, and post-training a larger model, where it reports early results showing greater gains.^[1][2]

Continue reading

• GGRPOGroup Relative Policy Optimization (GRPO) is a reinforcement learning algorithm for fine-tuning large language models that eliminates the separate critic (value) network used by PPOMentioned 6 times above
• NNemotron 3.5 LightningNVIDIA Nemotron 3.5 Lightning is an open-weights 30 billion parameter mixture-of-experts language model with 3 billion active parameters per token, released by NVIDIA on August 11Mentioned 3 times above
• DDeep ResearchDeep Research is a category of AI agent products that perform multi-step web research to produce comprehensive, citation-backed analytical reportsMentioned twice above
• FFinanceBenchFinanceBench is an AI benchmark for open-book financial question answering, designed to test whether large language models can answer the kinds of questions a financial analyst asks about a publicly traded…Listed in See also
• RogoRogo is a New York City based artificial intelligence company that builds an agentic AI platform purpose-built for the financial services industry.Listed in See also
• AAI in financeAI in finance is the use of artificial intelligence and machine learning across financial services to automate decisions and analyze data at scale.Listed in See also

More from See also

• Knowledge Distillation
• Small language model
• Hebbia

More to read

• Google Brain
• Artificial Intelligence
• Meta AI

Topics:AI AgentsAI BenchmarksAI CompaniesFinance AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23O. Ozan Koyluoglu, Richard Diehl Martinez, Thejas Venkatesh, Yuhao Zhang and Ashwin Paranjape, "Samaya Sterling: A post-trained model for state-of-the-art financial research", Samaya AI research blog, 28 October 2026. samaya.ai/...rl-post-trained-agentic-model
2. ^1 ^2 ^3 ^4Maithra Raghu (@maithra_raghu), "Introducing Samaya Sterling...", X, 28 October 2026. x.com/...2115488643477590241
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19Jeremy Kahn, "Samaya AI, startup building AI for financial services, raises $43.5 million in VC funding", Fortune, 14 May 2025. fortune.com/...ices-raises-43-5-million-vc-funding
4. ^1 ^2Tiffany Luck, "Samaya AI: Supercharging Financial Services", NEA blog, 14 May 2025. nea.com/...aya-ai-supercharging-financial-services
5. ^1 ^2"Samaya AI lands $43.5m to scale expert AI agents for finance sector", FinTech Global, 15 May 2025. fintech.global/...ert-ai-agents-for-finance-sector
6. ^"Samaya AI: $43.5 Million Raised To Advance AI Platform For Financial Services", Pulse 2.0, 18 May 2025. pulse2.com/...e-ai-platform-for-financial-services
7. ^1 ^2 ^3 ^4Maithra Raghu, "Samaya raises Series A led by NEA with $43.5 million in funding", Samaya AI blog. samaya.ai/...samaya-raises-series-a-led-by-nea
8. ^1 ^2 ^3 ^4 ^5Amrita Khalid, "Maithra Raghu", TIME100 AI 2025, TIME, 26 August 2025. time.com/...maithra-raghu
9. ^1 ^2"Two computer science alumnae among TIME100 AI list", Cornell Ann S. Bowers College of Computing and Information Science. cs.cornell.edu/...ce-alumnae-among-time100-ai-list
10. ^1 ^2Maithra Raghu, personal website. maithraraghu.com
11. ^1 ^2"Our founding team", Samaya AI, accessed 30 October 2026. samaya.ai/team
12. ^"Fabio Petroni, Group Leader", European Molecular Biology Laboratory (EMBL). embl.org/...fabio-petroni
13. ^1 ^2 ^3 ^4Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang, "Lost in the Middle: How Language Models Use Long Contexts", arXiv:2307.03172, 2023; Transactions of the Association for Computational Linguistics. arxiv.org/...2307.03172
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Samaya AI launches the Agent Control Plane (ACP); Welcomes NVentures (NVIDIA's venture capital arm) and Databricks Ventures as new investors", Samaya AI blog, 10 February 2026. samaya.ai/...samaya-ai-blog-acp
15. ^1 ^2 ^3 ^4"FinTech Samaya AI wins backing from NVentures", FinTech Global, 11 February 2026. fintech.global/...a-ai-wins-backing-from-nventures
16. ^1 ^2"Samaya AI Launches Agent Control Plane And Adds NVentures And Databricks Ventures Backing", Pulse 2.0, 11 February 2026. pulse2.com/...ures-and-databricks-ventures-backing
17. ^"Investors", Samaya AI, accessed 30 October 2026. samaya.ai/investors
18. ^1 ^2 ^3"Morgan Stanley customer story", Samaya AI. samaya.ai/...morgan-stanley
19. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Releases", Samaya AI, accessed 30 October 2026. samaya.ai/releases
20. ^1 ^2 ^3"Samaya MCP", Samaya AI. samaya.ai/...mcp
21. ^1 ^2"Research papers", Samaya AI, accessed 30 October 2026. samaya.ai/papers
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Yuhao Zhang, O. Ozan Koyluoglu, Thejas Venkatesh, Richard Diehl Martinez, Vishank Bhatia, Arash Alidoust and Ashwin Paranjape, "FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents", arXiv:2608.11683, 12 August 2026. arxiv.org/...2608.11683
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"Introducing FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents", Samaya AI blog, 9 July 2026. samaya.ai/...frontier-finance
24. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"FrontierFinance benchmark" (leaderboard updated 28 October 2026), Samaya AI. samaya.ai/...frontier-finance
25. ^"samaya-ai/FrontierFinance" dataset, Hugging Face. huggingface.co/...FrontierFinance
26. ^"samaya-ai/frontier-finance" (FrontierFinance benchmark grader), GitHub. github.com/...frontier-finance
27. ^1 ^2StepFun, "Step 5 Preview" model page, 20 September 2026. stepfun.com/step-5-preview
28. ^Arash Alidoust, Ashwin Paranjape and Maithra Raghu, "Frontier AI models outperform human experts on earnings prediction for the first time", Samaya AI blog, 8 October 2026. samaya.ai/...-human-experts-on-earnings-prediction
29. ^1 ^2Howard Yen, Yoonsang Lee, Ashwin Paranjape, Mengzhou Xia, Thejas Venkatesh, Jack Hessel, Danqi Chen and Yuhao Zhang, "Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search", arXiv:2510.18939, 2025. arxiv.org/...2510.18939
30. ^1 ^2Orion Weller, Benjamin Van Durme, Dawn Lawrie, Ashwin Paranjape, Yuhao Zhang and Jack Hessel, "Promptriever: Instruction-Trained Retrievers Can Be Prompted Like Language Models", arXiv:2409.11136, 2024. arxiv.org/...2409.11136
31. ^Skyler Hallinan, Thejas Venkatesh, Xiang Ren, Sai Praneeth Karimireddy, Ashwin Paranjape, Yuhao Zhang and Jack Hessel, "OpaqueToolsBench: Learning Nuances of Tool Behavior Through Interaction", arXiv:2602.15197, 2026. arxiv.org/...2602.15197
32. ^1 ^2 ^3Samaya AI, "Terms of Use" and "API" pages, accessed 30 October 2026. samaya.ai/terms ; samaya.ai/...api
33. ^1 ^2"Samaya blog" (index of company and research posts), Samaya AI, accessed 30 October 2026. samaya.ai/news

Show all 33 references