APEX-Agents: Investment Banking Analyst measures whether AI agents can execute real investment banking analyst tasks.
Across tasks graded by experts, the Gemini 3.7 Flash agent leads on the Loop harness at 77.1% Mean Score, ahead of the Opus 5.5 agent (73.7%) and Opus 5 agent (70.8%).
The table below shows the Top 10 AI models by Mean Score on the APEX-Agents: Investment Banking Analyst leaderboard using the Loop harness. Scores are as of September 2026.
| Rank | Agent | Mean Score (ReAct) |
|---|---|---|
| 1 | Gemini 3.7 Flash (High) | 77.1% |
| 2 | Opus 5.5 (Max) | 73.7% |
| 3 | Opus 5 (Max) | 70.8% |
| 4 | Grok 4.6 (xHigh) | 70.3% |
| 5 | Fable 5 (Max) | 69.3% |
| 6 | Qwen3.8-Max (xHigh) | 69.1% |
| 7 | Fable 5.1 (Max) | 66.7% |
| 8 | GLM-5.3 (Max) | 64.6% |
| 9 | DeepSeek-V4-Flash (Max) | 61.9% |
| 10 | Opus 4.8 (Max) | 61.5% |
Scores: Mean Score on the Loop harness (loop_truncated_tools_agent).
View full APEX-Agents: Investment Banking Analyst leaderboard →
How to read the APEX-Agents: Investment Banking Analyst leaderboard?
APEX-Agents evaluates each agent for investment banking analyst tasks on the Loop harness, using 2 scoring methods.
Harness architecture:
- Loop architecture: the agent operates in a truncated reasoning cycle, testing whether agents can complete structured financial modeling without extended step-by-step planning.
Scoring:
- Pass@1 Score: is calculated on whether the agent completes the full task correctly on its 1st attempt. For the sample merger model task, all 8 sensitivity outputs must be correct simultaneously.
- Mean Score: gives partial credit for each rubric criterion met. A merger model with 5 of 8 sensitivity cells correct still earns credit for what it got right.
Top 5 Investment Banking Analyst AI Agent Models by Loop, Pass@1:
- Gemini 3.7 Flash (High) 71.3%
- Opus 5.5 (Max) 69.3%
- Opus 5 (Max) 66.6%
- Grok 4.6 (xHigh) 65.0%
- Qwen3.8-Max (xHigh) 64.2%
View full leaderboard (Loop, Pass@1) →
Top 5 Investment Banking Analyst AI Agent Models by Loop, Mean Score:
- Gemini 3.7 Flash (High) 77.1%
- Opus 5.5 (Max) 73.7%
- Opus 5 (Max) 70.8%
- Grok 4.6 (xHigh) 70.3%
- Fable 5 (Max) 69.3%
View full leaderboard (Loop, Mean Score) →
Who should use APEX-Agents: Investment Banking Analyst?
APEX-Agents: Investment Banking Analyst serves 3 types of users:
- Individual IB analysts and associates evaluating which AI agent handles financial modeling tasks most reliably for their deal work.
- Banking platform and technology teams standardizing AI agent deployments across deal teams. The benchmark gives them expert-graded performance data on tasks analysts actually do.
- AI labs and companies building IB-specific agentic products that need expert-graded performance data on real financial tasks.
How are AI model rankings evaluated for investment banking analysts?
These rankings come from APEX-Agents: Investment Banking Analyst, Mercor's benchmark for how frontier AI agents perform on real IB analyst tasks. Traditional AI benchmarks measure general reasoning or isolated calculations. APEX-Agents was built to score the multi-step, tool-assisted work IB analysts actually do: merger modeling, accretion/dilution analysis, and sensitivity tables built to specific deal assumptions.
What does the APEX-Agents: Investment Banking Analyst benchmark measure?
APEX-Agents: Investment Banking Analyst measures whether an AI agent running on frontier AI models can perform economically valuable IB analyst tasks across 160 cases. Each task requires the agent to operate across multiple applications, reason over multiple steps, and produce outputs with the precision deal teams require. Tasks are graded by human-authored rubrics with exact numeric answer requirements. Built with practitioners from Goldman Sachs, Morgan Stanley, JPMorgan, and Barclays.
How does APEX-Agents: Investment Banking Analyst benchmark scoring work?
Every rubric is written by an experienced IB professional who defines what correct means for that task. Each criterion is a binary yes/no graded against a precise expected output. For the sample accretion/dilution task, one criterion reads: States that BBDC accretion/dilution is 36.83% for 10.0% Bid Premium and 15.0% Cash Consideration. In financial modeling, the difference between 36.83% and any other figure is not acceptable variation, it is a wrong number that would not survive deal team review.
What type of task is evaluated by the APEX-Agents: Investment Banking Analyst benchmark?
Here's one of the sample tasks that was evaluated within this domain:
Sample task: M&A merger model, accretion/dilution sensitivity analysis
The agent receives an existing merger model and must edit it to produce two sensitivity tables: one for BBDC shareholder accretion/dilution and one for TVPG shareholder accretion/dilution, each sensitized to bid premium (10% and 20%) and cash consideration (10% and 15%).
The agent must apply a 480bps increase in EBIT synergies and a 210bps decrease in post-deal bidder share price downside, and output all values as percentages rounded to two decimal places.
The rubric has 8 criteria, each checking one output cell against an exact value. An agent that gets most sensitivity cells right but misses one still earns partial credit under Mean Score while failing the task under Pass@1.
What are the hardest investment banking tasks for AI agents?
Multi-variable sensitivity analysis and structured financial modeling remain the hardest tasks on APEX-Agents: Investment Banking Analyst. The top-ranked agent scores 77.1% on Loop Mean Score, but performance drops off sharply: some models such as GPT-OSS-120B score 0.9% on Loop Pass@1, showing that strong general AI capability does not translate to precise financial modeling.
- Exact numeric outputs are required. An agent that reasons correctly about accretion/dilution but produces a rounded figure still fails the rubric criterion.
- Multi-variable sensitivity analysis compounds errors. The sample task requires 8 distinct outputs across combinations of two variables for two shareholder groups.
- Deal assumptions must be applied precisely. Adjusting EBIT synergies by exactly 480bps while preserving all other model relationships requires instruction-following that general agents frequently fail.
- Senior analysts still outperform AI agents on precision deliverables. Structured merger modeling and exact sensitivity analysis remain largely the preserve of human expertise.
AI agents accelerate IB work on drafting, research, and template preparation. They do not yet reliably complete the precision financial modeling that deal teams depend on without expert review.
Do AI models or AI tools for investment banking matter more?
Both matter, but they play different roles. The underlying model like Gemini 3.7 Flash, Opus 5.5, or Opus 5 determines the agent's reasoning capability and sets the performance ceiling. The tool like an Excel add-in, a financial data platform, or a Copilot integration wraps that capability in context and delivers it into analyst workflows.
A well-integrated tool cannot compensate for a weak underlying model on tasks requiring financial precision. No Excel workflow closes a 20-point Pass@1 gap on structured financial modeling. However, a good tool amplifies what the underlying model offers, giving the agent access to deal data, model templates, and financial formatting that a raw API call does not have.
For research summarization and narrative drafting, the difference between the top models is smaller and your tool and workflow often decide the result.
For the precision-intensive work APEX-Agents measures, structured merger modeling and exact sensitivity analysis, the underlying model's capability is the deciding factor. Pick the model for the ceiling you need; pick the tool for how effectively it delivers that capability into your deal team's workflow.
Get the full APEX-Agents: Investment Banking Analyst dataset
The public leaderboard shows the scores. The full dataset is what banking technology teams and AI labs use to make real deployment decisions, including task-level breakdowns, rubric details, and agent trajectories across all 160 tasks.
For teams that need more
Evaluate AI agent performance on your own investment banking workflows with custom evaluations graded by Mercor's network of IB professionals.
Frequently Asked Questions
What is the best AI agent for investment banking analyst tasks right now?+−
Gemini 3.7 Flash agent leads APEX-Agents: Investment Banking Analyst at 77.1% Mean Score on the Loop harness, ahead of the Opus 5.5 agent (73.7%) and Opus 5 agent (70.8%). On Pass@1, Gemini 3.7 Flash also leads at 71.3%, with Opus 5.5 second at 69.3% and Opus 5 third at 66.6%. Rankings shift with each new model release.
Will AI replace investment banking analysts?+−
The top-ranked agent scores 77.1% Mean Score on APEX-Agents: Investment Banking Analyst tasks on the Loop harness. On Pass@1, the top agent completes 71.3% of tasks on its first attempt, while many models score far lower. AI agents accelerate IB work on research, drafting, and data preparation. They do not reliably perform the precision financial modeling that deal teams depend on.
How often is the APEX-Agents leaderboard for Investment Banking Analyst updated?+−
The leaderboard is updated as new frontier models are evaluated. Rankings shift with each major release. Verify scores at the time of your decision on the live leaderboard.
How can my team evaluate AI agents on our own deal workflows?+−
Mercor offers custom evaluations graded by vetted IB professionals on firm-specific financial modeling tasks, giving banking technology teams a benchmark built on their own workflows. Get in touch.
What AI tools do investment banking analysts use?+−
Microsoft Copilot integrated into Excel is common at firms with Microsoft 365 enterprise licenses. Shortcut is a purpose-built Excel add-in for financial modeling that has been benchmarked against general-purpose models in third-party evaluations. Bloomberg and CapIQ both offer AI-assisted features for data extraction and analysis within their platforms.
Can AI agents build financial models?+−
AI agents can edit and extend existing financial models, but reliably building a complex model from scratch remains beyond current agents. The sample task on APEX-Agents: Investment Banking Analyst provides an existing merger model for the agent to edit. Even with that head start, the top-ranked agent completes 71.3% of tasks on its first attempt on the Loop harness, Pass@1, leaving a meaningful error rate. Output from AI agents on financial modeling requires expert review before it reaches a deal team.
Can I use ChatGPT or Claude on confidential deal materials?+−
Enterprise tiers of most frontier model providers offer zero-data-retention options that do not use inputs to train future models. Before using any AI agent on live deal data, confirm the specific data-handling terms for your tier and check whether your firm's compliance policies permit it.