General-purpose AI leaderboards can tell us a great deal about a model’s reasoning, language, and broad knowledge capabilities. What they do not tell us is how effectively that model can answer real, institutional-grade financial research questions.
For investment teams, research providers, and financial technology platforms, the more relevant question is narrower:
When connected to professional-grade financial data, which models produce the most accurate, complete, and source-grounded answers to the questions an analyst would actually ask?
The Aiera Leaderboard was designed to answer that question.
Rather than treating financial research as one task among many, the Aiera Leaderboard makes it the primary basis for evaluating and ranking large language models. The result is a research-gated, research-weighted benchmark that produces a meaningfully different view of model performance.
Why Financial Research Requires a Different Benchmark
Many established AI evaluations focus on general reasoning, coding, mathematics, factual recall, or performance across broad public datasets. Financial benchmarks often test numerical reasoning or summarization when the relevant source document has already been provided.
Those evaluations are useful, but they do not fully reflect an institutional research workflow.
In practice, an AI system must often:
- Identify and retrieve the right source
- Work across research, expert insights, transcripts, filings, and other content types
- Distinguish relevant facts from surrounding information
- Synthesize information across multiple documents
- Produce an answer that is accurate, complete, and grounded in source material
This requires more than general model capability. It requires the ability to use professional financial data effectively.
How the Aiera Score Works
The Aiera Score combines four financial tasks, each scored on a 0 to 100 scale:
The leaderboard is research-gated, meaning a model must complete the Research evaluation to be listed.
Research receives 60% of the total weight because producing correct, sourced answers from institutional content is the most important capability being measured. The remaining 40% captures established financial language and reasoning tasks.
Within that capability component, question answering carries the greatest weight, followed by summarization and sentiment analysis.
Measuring Research Performance
For the Research task, each model is connected to the same Aiera MCP (Model Context Protocol) server and evaluated against proprietary, analyst-grade financial questions.
A held-out judge evaluates the answer at the individual fact level, measuring the share of required key facts that the model successfully supports.
The Research score represents the model’s absolute performance while connected to Aiera. It does not measure how much the model improved relative to a version without Aiera access.
That distinction is important.
The Aiera Leaderboard ranks models based on how well they ultimately perform. The separate Aiera Lift study measures the improvement created by connecting a model to Aiera’s financial data.
The Current Aiera Leaderboard
The June 2026 evaluation includes 16 research-qualified models.
Research Performance Is a Distinct Capability
The most important finding is not simply which model ranks first. It is that research performance is only weakly related to general financial capability.
Across the 16 models, the correlation between Research performance and the capability-only score was approximately 0.34.
That means a model’s strength in financial question answering, summarization, and sentiment analysis does not reliably predict how well it will answer research questions when connected to external data.
Only five of the 16 models scored above 30 on Research, while seven scored below 10.
General capability helped order the field, but Research performance separated the leaders.
The Top Rankings Are Not an Artifact of the Weighting
Giving Research 60% of the Aiera Score is a deliberate reflection of its importance to institutional workflows. The weighting was also tested to determine whether small changes would materially alter the results.
They did not.
When the Research weight was varied from 40% to 70%, the same four models remained at the top, with Claude-Opus-4.7 and GPT-5.5 exchanging third and fourth place at some settings, but the group itself remained unchanged.
Composite Scores Do Not Tell the Entire Story
The individual task scores reveal important differences in how models behave.
Kimi-K2.6 ranks sixth overall and posts strong Research and financial question-answering scores of 31.6 and 86. At the same time, its Summary and Sentiment scores are only 9 and 11, reflecting empty or malformed outputs on those tasks.
Qwen3-235B presents the opposite type of result. It records a strong Q&A score of 84, along with solid Summary and Sentiment performance, but achieves a Research score of only 2.6.
Gemini-2.5-Flash performs consistently across the general capability tasks but records a more modest Research score of 12.5.
These profiles reinforce the need to evaluate both the composite and the underlying tasks. The best model for one workflow may not be the best model for another.
What the Results Mean for Financial Institutions
The findings have several practical implications.
Model selection should reflect the actual workflow
A model chosen for document summarization may not be the best model for open-ended research across a large content library. Financial institutions should test models against the tasks, data, and delivery methods they expect to use in production.
Data access alone is not sufficient
Connecting a model to high-quality financial content creates the opportunity for better answers, but the model must still retrieve, interpret, and synthesize that information effectively.
General leaderboards should not be treated as financial research rankings
Broad model evaluations remain valuable, but they answer a different question. Institutional research requires its own benchmark, grounded in representative questions and professional content.
Per-task transparency matters
A single composite score is useful for ranking, but individual task results reveal limitations that could materially affect a production deployment.
A More Relevant Way to Evaluate Financial AI
The Aiera Leaderboard is designed around a simple principle: financial AI should be evaluated on the work financial professionals need it to perform.
By making grounded research the primary metric, the leaderboard surfaces differences that general-purpose benchmarks miss. It shows that research performance is a distinct and highly discriminating capability, that leading frontier models currently hold a clear advantage, and that several open-weight models remain competitive in specific areas.
As financial institutions bring AI into more research workflows, model evaluation must move beyond broad measures of intelligence and toward the quality of the answers produced from the data that actually matters.
Download the full Aiera Leaderboard paper to review the complete rankings, methodology, task-level results, and limitations.
Introducing the Aiera Leaderboard: A Research-Centric Benchmark for Financial LLMs
General-purpose AI leaderboards can tell us a great deal about a model’s reasoning, language, and broad knowledge capabilities. What they do not tell us is how effectively that model can answer real, institutional-grade financial research questions.
For investment teams, research providers, and financial technology platforms, the more relevant question is narrower:
When connected to professional-grade financial data, which models produce the most accurate, complete, and source-grounded answers to the questions an analyst would actually ask?
The Aiera Leaderboard was designed to answer that question.
Rather than treating financial research as one task among many, the Aiera Leaderboard makes it the primary basis for evaluating and ranking large language models. The result is a research-gated, research-weighted benchmark that produces a meaningfully different view of model performance.
Why Financial Research Requires a Different Benchmark
Many established AI evaluations focus on general reasoning, coding, mathematics, factual recall, or performance across broad public datasets. Financial benchmarks often test numerical reasoning or summarization when the relevant source document has already been provided.
Those evaluations are useful, but they do not fully reflect an institutional research workflow.
In practice, an AI system must often:
This requires more than general model capability. It requires the ability to use professional financial data effectively.
How the Aiera Score Works
The Aiera Score combines four financial tasks, each scored on a 0 to 100 scale:
The leaderboard is research-gated, meaning a model must complete the Research evaluation to be listed.
Research receives 60% of the total weight because producing correct, sourced answers from institutional content is the most important capability being measured. The remaining 40% captures established financial language and reasoning tasks.
Within that capability component, question answering carries the greatest weight, followed by summarization and sentiment analysis.
Measuring Research Performance
For the Research task, each model is connected to the same Aiera MCP (Model Context Protocol) server and evaluated against proprietary, analyst-grade financial questions.
A held-out judge evaluates the answer at the individual fact level, measuring the share of required key facts that the model successfully supports.
The Research score represents the model’s absolute performance while connected to Aiera. It does not measure how much the model improved relative to a version without Aiera access.
That distinction is important.
The Aiera Leaderboard ranks models based on how well they ultimately perform. The separate Aiera Lift study measures the improvement created by connecting a model to Aiera’s financial data.
The Current Aiera Leaderboard
The June 2026 evaluation includes 16 research-qualified models.
Research Performance Is a Distinct Capability
The most important finding is not simply which model ranks first. It is that research performance is only weakly related to general financial capability.
Across the 16 models, the correlation between Research performance and the capability-only score was approximately 0.34.
That means a model’s strength in financial question answering, summarization, and sentiment analysis does not reliably predict how well it will answer research questions when connected to external data.
Only five of the 16 models scored above 30 on Research, while seven scored below 10.
General capability helped order the field, but Research performance separated the leaders.
The Top Rankings Are Not an Artifact of the Weighting
Giving Research 60% of the Aiera Score is a deliberate reflection of its importance to institutional workflows. The weighting was also tested to determine whether small changes would materially alter the results.
They did not.
When the Research weight was varied from 40% to 70%, the same four models remained at the top, with Claude-Opus-4.7 and GPT-5.5 exchanging third and fourth place at some settings, but the group itself remained unchanged.
Composite Scores Do Not Tell the Entire Story
The individual task scores reveal important differences in how models behave.
Kimi-K2.6 ranks sixth overall and posts strong Research and financial question-answering scores of 31.6 and 86. At the same time, its Summary and Sentiment scores are only 9 and 11, reflecting empty or malformed outputs on those tasks.
Qwen3-235B presents the opposite type of result. It records a strong Q&A score of 84, along with solid Summary and Sentiment performance, but achieves a Research score of only 2.6.
Gemini-2.5-Flash performs consistently across the general capability tasks but records a more modest Research score of 12.5.
These profiles reinforce the need to evaluate both the composite and the underlying tasks. The best model for one workflow may not be the best model for another.
What the Results Mean for Financial Institutions
The findings have several practical implications.
Model selection should reflect the actual workflow
A model chosen for document summarization may not be the best model for open-ended research across a large content library. Financial institutions should test models against the tasks, data, and delivery methods they expect to use in production.
Data access alone is not sufficient
Connecting a model to high-quality financial content creates the opportunity for better answers, but the model must still retrieve, interpret, and synthesize that information effectively.
General leaderboards should not be treated as financial research rankings
Broad model evaluations remain valuable, but they answer a different question. Institutional research requires its own benchmark, grounded in representative questions and professional content.
Per-task transparency matters
A single composite score is useful for ranking, but individual task results reveal limitations that could materially affect a production deployment.
A More Relevant Way to Evaluate Financial AI
The Aiera Leaderboard is designed around a simple principle: financial AI should be evaluated on the work financial professionals need it to perform.
By making grounded research the primary metric, the leaderboard surfaces differences that general-purpose benchmarks miss. It shows that research performance is a distinct and highly discriminating capability, that leading frontier models currently hold a clear advantage, and that several open-weight models remain competitive in specific areas.
As financial institutions bring AI into more research workflows, model evaluation must move beyond broad measures of intelligence and toward the quality of the answers produced from the data that actually matters.
Download the full Aiera Leaderboard paper to review the complete rankings, methodology, task-level results, and limitations.