How OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI

Highlights

• OpenAI launched ChatGPT for Financial Services as a tailored product aimed at investment banking and equity research, developed with design partners Morgan Stanley and Evercore.

• OpenAI solved previous pre-training scaling barriers in 2026, training larger and more capable models like GPT-5.6 Sol and GPT-6 Astra.

• Internal OpenAI models proposed a solution to the Navier-Stokes millennium prize problem and shared 719 mathematical research manuscripts across 372 families of results.

• By the end of 2027, models are expected to outperform leading human experts on a growing range of difficult tasks across different fields and specialties.

• OpenAI plans to invest heavily in finance-specific post-training over a three-month period to improve general financial reasoning, Excel model generation, and PowerPoint presentation creation.

Timeline

1. 00:00 - 01:26

   Introduction to ChatGPT for Financial Services

   Joseph Kim introduces the new product designed to bring AI-driven productivity changes seen in software engineering over to financial industries. The development process involved close collaboration with design partners Morgan Stanley and Evercore to address critical industry requirements such as native data access, familiar financial interfaces, and custom formatting for templates.

   • ChatGPT for Financial Services launched one month prior as a tailored version of ChatGPT work focused on investment banking and equity research.
   • Product development partnered with major financial institutions Morgan Stanley and Evercore.
   • The core focus areas include financial data integration, an interface built for finance, and adherence to firm-specific formats.
2. 04:12 - 05:58

   Addressing the Capability Overhang

   Model intelligence has scaled rapidly from high-school math problems to solving advanced challenges, yet everyday workflows have remained static for many professionals. Examining internal company data reveals that product form factor adjustments, such as dedicated desktop applications, act as major inflection points that drive adoption outside of software engineering.

   • The gap between rapid model capability improvements and lagging day-to-day work utilization is defined as the capability overhang.
   • Codex usage across non-engineering departments at OpenAI surged from nearly zero to about 90% following the release of the Codex desktop app in February 2026.
   • Bridging the capability overhang requires specialized user experiences, appropriate tools, and relevant data rather than just capable base models.
3. 05:58 - 09:18

   Technical Lessons in Data and Architecture

   Initial evaluations of third-party MCP connectors revealed latency and search quality issues, prompting direct data licensing and dedicated hosting. While MCP and skills provide external tool access and instructions, post-training embeds structural familiarity directly into the model weights. Additional features like citation viewers and interactive charts facilitate data verification within existing workflows.

   • OpenAI licensed financial data directly and built custom hosting and search indices on web index infrastructure to ensure high reliability and low latency.
   • Model Context Protocol (MCP) and skills function like instruction manuals for junior analysts, whereas post-training acts like a senior analyst with decades of experience.
   • A specialized template system enables models to generate spreadsheets and presentations that match organizational formats automatically.
4. 09:18 - 15:00

   Predictability of Model Progress and Future Capabilities

   Historical development shows predictable trends driven by scaling laws and algorithmic refinement. Breakthroughs in reasoning via reinforcement learning have unlocked advanced problem-solving, moving from standardized testing to producing original research in mathematics. The threshold phenomenon dictates that practical adoption accelerates sharply only after verification time drops below the time saved by the tool.

   • Model progress is driven by scaling compute, improving algorithms, and unhobbling techniques, with training compute dropping by half roughly every eight months.
   • Synthetic data generation, verification rewards, and sparse attention have successfully overcome anticipated data walls and context window limits.
   • The threshold phenomenon explains why steady underlying model progress feels sudden once performance crosses the boundary of practical usefulness.
5. 15:00 - 18:52

   Future Projections and Bets for Finance in 2027

   General capabilities enhanced by reasoning improvements combine with targeted financial post-training to elevate AI performance on complex workflows like financial analysis and forecasting. Transitioning from reactive chat interfaces to persistent, goal-oriented agents will fundamentally alter professional workflows. By 2027, AI models are anticipated to execute sophisticated economic judgments and match or exceed expert performance levels.

   • Progress in finance relies on a rising tide of general model improvements and specific waves driven by finance-focused post-training.
   • Un-hobbling models through proactive and persistent agents will allow AI to hold goals and execute background tasks without continuous prompting.
   • Three explicit bets for 2027 include the elimination of manual Excel and PowerPoint creation, human-like co-worker agent dynamics, and models outperforming human experts in forecasting.