AI Engineer hosted Joseph Kim, who leads OpenAI's ChatGPT for Financial Services, a ChatGPT Work variant Kim said launched a month ago for investment banking and equity research.
Kim said internal Codex output for finance, legal and recruiting rose from near zero in early 2026 to about 90 percent by June after the February desktop app, while engineering use grew steadily.
Kim said compute needed for a given language model performance has fallen by roughly half every eight months, a steady gain he tied to scaling, algorithms and unhobbling.
Kim said GPT-4 scored 700 out of 800 on SAT math in 2023, while an internal model by early 2026 disproved a 1946 conjecture and shared 719 manuscripts across 372 families.
Per Tipalti, the finance product ships with licensed data from LSEG, PitchBook and Daloopa, which contrasts with the consumer Finances in ChatGPT flow where Plaid connects personal accounts according to OpenAI's help center.
Kim did not disclose paid seat counts, pricing, or audited error rates on generated Excel models and decks, so the share and capability curves do not show accuracy or retention in live deals.
Consistent audit pass rates and earnings forecast benchmarks would change the read, and the open question is whether finance post-training plus the rising tide pushes peaks past leading human experts by late 2027?
