Ozan Koyluoglu retweeted
Introducing Samaya Sterling: our first post-trained model for state-of-the-art financial research.
On FrontierFinance, Sterling outperforms Claude Fable 5.1 (58.9% vs. 55.9%) at 11× lower cost.
Using a model with just 3B active parameters, Sterling demonstrates what's possible when specialized training meets a purpose-built financial research harness.
Read the full research ↓
New research at @samaya_AI : Hindsight Structured Advice Distillation (HSAD). We distill hindsight from an agent's graded runs into a small advisor that writes structured advice for each question. On FrontierFinance: 52.9% → 56.6%. See blog for details. samaya.ai/blog/hindsight-str…
Ozan Koyluoglu retweeted
Excited to see Gemini 3.6 Flash deliver frontier-level accuracy at $2.41/query and 164s.
Try Gemini 3.6 Flash for financial use cases, and share your feedback! A great combination of speed, cost-efficiency, and high performance
New results in: Gemini 3.6 Flash achieves 46.3% on FrontierFinance, our open benchmark for finance AI agents. Ahead of Claude Opus 4.8, on par with GPT-5.6 Sol, and a big jump over past Gemini generations.
What stands out is the efficiency: it hits that score at just $2.41 per query, cheaper than both, and at 164s it's among the fastest models we've tested on this harness.
In our analysis, it's notably stronger than Opus 4.8 at surfacing qualitative and contextual insight, and leads on Screening & Discovery, one of the benchmark's hardest use cases.
More about FrontierFinance and the full benchmarking results 👇
Ozan Koyluoglu retweeted
New results in: Gemini 3.6 Flash achieves 46.3% on FrontierFinance, our open benchmark for finance AI agents. Ahead of Claude Opus 4.8, on par with GPT-5.6 Sol, and a big jump over past Gemini generations.
What stands out is the efficiency: it hits that score at just $2.41 per query, cheaper than both, and at 164s it's among the fastest models we've tested on this harness.
In our analysis, it's notably stronger than Opus 4.8 at surfacing qualitative and contextual insight, and leads on Screening & Discovery, one of the benchmark's hardest use cases.
More about FrontierFinance and the full benchmarking results 👇
Ozan Koyluoglu retweeted
𝗢𝘂𝗿 𝗻𝗲𝘄 𝗵𝗶𝗴𝗵𝗲𝗿-𝗲𝗳𝗳𝗼𝗿𝘁 𝘃𝗲𝗿𝘀𝗶𝗼𝗻 𝗼𝗳 𝗦𝗮𝗺𝗮𝘆𝗮 𝗮𝗰𝗵𝗶𝗲𝘃𝗲𝘀 𝟱𝟲% 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆, 𝗮 𝗻𝗲𝘄 𝘀𝘁𝗮𝘁𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗮𝗿𝘁, 𝗮𝘁 𝟮𝘅 𝗹𝗼𝘄𝗲𝗿 𝗰𝗼𝘀𝘁 𝘁𝗵𝗮𝗻 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲 𝟱.
We built FrontierFinance with a simple goal: improve AI for investment decision-making. After releasing the benchmark, our team identified opportunities to improve tool design, optimize context management, and increase reasoning effort.
These changes improve performance across Screening & Discovery, Financial Data Extraction, Earnings & Events, and Company Research, helping Samaya extract quantitative and qualitative data, along with improved contextual reasoning and analysis.
Full announcement below 👇
Ozan Koyluoglu retweeted
Excited to be releasing FrontierFinance, the largest and most challenging open benchmark for evaluating AI agents across the full investment workflow!
FrontierFinance is substantially harder than current finance benchmarks: Existing benchmarks like FinanceBench and Finance Agent focus almost entirely on data extraction.
FrontierFinance spans diverse use cases across the full investment process: Screening & Discovery, Company Research, Sector/Industry/Macro, Earnings & Events, and Coverage & Catalyst Monitoring.
Created for ambiguous, long-horizon agents: 220 examples paired with 11,543 expert-crafted rubrics, following Samaya's Criteria Eval methodology. The rubrics are what let us evaluate the reasoning and steps behind a true expert-level output, not just a plausible-looking one.
Evaluations: We evaluated Claude Fable 5, Claude Opus 4.8, GPT 5.5, Gemini, open-source models including GLM and DeepSeek, and others. We used the same public rubric and a standard harness for financial tasks. Samaya's AI system reached state-of-the-art accuracy at 50.8%, at 4x lower inference cost than Fable 5. Next best was Fable 5 (49.2%), then Opus 4.8 (45%) and GPT 5.5 (43.5%).
We're releasing the benchmark, methodology, and full evaluation results - see link in comments.
Future releases: FrontierFinance was curated from Samaya's larger internal set of ~5,000 examples, and we plan to release subsequent, harder benchmarks as well as a more detailed technical report!
We have just released FrontierFinance, the largest and most challenging open benchmark for evaluating ambiguous, long-horizon AI agents for finance workflows! Built from research at @samaya_AI .
7
35
Website: research.samaya.ai/benchmark…
Blog: samaya.ai/blog/frontier-fina…
10
We're excited to share the dataset, methodology, and code with the community today, and we plan to release a detailed technical report soon.
9
Leading models tested: Evaluated top frontier and open-source models. Claude Fable 5 hits 49.2%, Opus 4.8 at 45.0%, and GPT 5.5 at 43.5%. We also tested Gemini, GLM, and DeepSeek (with DeepSeek V4 Pro notably closing the gap at just 26% of the cost).
40
SOTA performance: Samaya's in-house AI system leads at 50.8%, outperforming the next best model at a 4x lower inference cost.
13
Deep evaluation: 220 queries paired with 11,543 expert-crafted rubrics. We use Samaya's Criteria Eval methodology to grade the reasoning and steps behind an agent's output, ensuring it meets professional standards.
12
Comprehensive workflows: Moves beyond simple extraction to cover real-world use cases: Screening & Discovery, Company Research, Sector/Industry/Macro, Earnings & Events, and Coverage & Catalyst Monitoring.
14
Unlike standard benchmarks that focus heavily on data extraction, FrontierFinance tests true expert-level reasoning and synthesis.
19
Ozan Koyluoglu retweeted
@Samaya_AI trained a query understanding model that outperforms Claude Sonnet 4.6 and Gemini Flash on financial queries, with lower latency.
Better query understanding means better agents.
Here's how we trained it and the results 👇
@AshwinParan @_vThejas @ozankoyluoglu @yuhaozhangx @richarddm1
Ozan Koyluoglu retweeted
Why does query understanding matter?
Financial research agents need to correctly identify critical attributes such as the company and reporting period before they can retrieve the right evidence.
We trained the model using a two-stage approach combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which led to:
🔹+1.5pp quality
🔹4% fewer tool calls
🔹23% lower median time to answer
🔗 Full blog: samaya.ai/blog/query-underst…
Ozan Koyluoglu retweeted
orta direk, beyaz yaka. Adı ne derse ne diyelim. 20 senedir ülkeyi yönetenlerin elinde paspas oldu. Üstten zenginler, alttan ise kasabalı cahiller arasında posası çıkana kadar sağıldı. Ne mecliste, ne sokakta hiçbir yerde temsiliyeti yok. Test emekçisi. Çocukluğundan beri test çözüyor. Ülkede vergisini aksatmayan tek grup. Zaten eline geçmeden alınıyor. Bodrum da tatile gitse öpülür, araba alsa öpülür. Yurtdışına çıksa harç öder, çocuğu ile kar tatiline gitse yanarak ölür. İstanbul da belediye kazanır. Başkanı hapse atılır. Taksici kendisi yerine Arap turist alır, turizmcisi Rus turist alır. Bizimkisi sırtında özel okul sistemini, özel hastaneleri, özel sektörü taşır. Akp en çok bu gruba zarar verdi. Zaten bunlar bana da oy vermiyor diyerek vurdukça vurdu. Bir itiraz gelmiyordu uzun zamandır. Boykotlar ile ilk öfke patlamaları başladı bana kalırsa.
Ozan Koyluoglu retweeted
Aziz milletim;
Asla umudunuzu yitirmeyin.
Göreceksiniz; haktan ve hukuktan nasibini almamış bir avuç insan kaybedecek, millet kazanacak.
Milyonlar bugün iradesine sahip çıktı.
Teşekkürler İstanbul.













