Singular Bank's 90-Minute Productivity Hack Cost $50M in AI Bills

Last week I watched a junior analyst at Goldman manually format portfolio slides for three hours. Same spreadsheet, different fonts, endless tweaking. Meanwhile, across town at Singular Bank, their Singularity assistant was cranking out the same work in minutes.

Singular Bank—that $12B AUM neobank everyone forgot about until now—just dropped their internal AI numbers. They're saving 60-90 minutes per banker daily across 500+ staff using OpenAI's ChatGPT and Codex. The breakdown:

• Meeting prep: Auto-generated agendas from emails and client data
• Portfolio analysis: Real-time risk assessments and scenario modeling
• Follow-ups: Personalized emails and compliance checklists

Built in six weeks. 95% accuracy on portfolio summaries. Sounds too good to be true.

The Expensive Engine Under the Hood

Here's where it gets spicy. Singularity runs on GPT-5.5 for the heavy reasoning work—portfolio modeling, risk analysis, the stuff that actually matters. At $5 per million input tokens and $30 per million output tokens, those 90 daily minutes aren't cheap.

Do the math: 500 bankers × 90 minutes × 250 working days. That's 18.75 million saved minutes annually. Sounds impressive until you calculate the token burn.

Sure, but what about the ongoing costs? A typical portfolio analysis might churn through 50K tokens of financial data input and generate 20K tokens of analysis. At enterprise volumes, we're talking serious money.

Multi-Agent Madness

Singular's setup leverages Codex's February 2026 multi-agent update—separate threads handling data extraction while others generate visuals. One agent pulls CRM data from Salesforce, another crunches numbers in their Aladdin integration, a third formats everything for Outlook.

It's elegant. It's also complex as hell.

The real story here isn't the productivity gains—it's the infrastructure bet. Singular partnered with TCS (one of OpenAI's enterprise scaling partners) to build this thing. They're running GPT-5.4 for basic tasks, GPT-5.5 for the complex stuff, and GPT-5.3-Codex-Spark for low-latency financial calculations.

That's three different model endpoints, custom fine-tuning on bank-specific prompts, and integration with legacy banking systems that probably still run on COBOL somewhere.

The Neobank Advantage

What Singular pulled off here would be impossible at JPMorgan or Goldman. Not technically—financially and politically. Legacy banks have:

1. Compliance departments that move like glaciers

2. IT procurement cycles measured in quarters

3. Risk committees allergic to "experimental" AI

Singular raised $450M in 2024 specifically to build "AI-native banking." When you're a startup with Sequoia and SoftBank money, you can afford to burn tokens on productivity experiments.

The timing matters too. They got early access to Codex's "Omnipotent Assistant" update in January 2026, right before the multi-agent features launched. First-mover advantage in the AI productivity race.

Following the Money

Let's be cynical about the metrics. 60-90 minutes saved daily sounds great, but what are those bankers doing with the extra time? Singular claims it's "high-value activities"—but I've seen productivity tools just create space for more meetings.

The real test: revenue per banker. If Singularity actually moves the needle on deal flow or client satisfaction, other banks will copy this overnight. If it just makes slide decks prettier, it's expensive window dressing.

Alexander Embiricos from OpenAI's Codex team predicts "10x productivity in knowledge work." Bold claim. Banking isn't software development—you can't just ship faster and iterate.

My Bet: Singular's productivity gains are real, but the competitive advantage evaporates within 18 months. Every major bank will have similar tools by 2027. The winners won't be the first adopters—they'll be whoever figures out the unit economics first.