Over the past few weeks, we’ve been exploring methods to extract structured data from unstructured financial documents, such as income statements, balance sheets, and cash flow statements. Our initial prototypes leveraged GPT-4o and GPT-4.1. We recently evaluated GPT-5 to assess its comparative performance for this use case.

Key Findings:

1. Processing Speed: GPT-5 demonstrates notably higher latency, with identical requests requiring 45-60 seconds compared to 10-15 seconds on GPT-4o.
2. Extraction Accuracy: GPT-5 delivers superior consistency and accuracy, successfully extracting all 18 target parameters from PDF documents, while GPT-4o typically captures only 14-16 parameters.
3. Interpretability: GPT-5 provides enhanced transparency in its extraction logic, including page references, operational definitions, and detailed reasoning paths that facilitate human understanding of the extraction process.

Our initial assessment: GPT-5 shows considerable improvement from older models for financial document processing applications where accuracy and interpretability take precedence over response time.

Under the Hood

Designing Complex Data Tables for Mobile

Designing complex data tables for mobile isn’t about shrinking desktop tables! Our product designer shared practical strategies for transforming complex tables into mobile friendly formats without sacrificing usability.

Getting Started With Swift Testing

Outlining a clear and expressive approach to testing in Swift - something beyond syntax improvements! How Swift delivers advanced capabilities including native parameterized testing with automatic parallel execution, and much more.

Breakroom Buzz

Introducing GPT-5

OpenAI has just unveiled GPT-5, it’s most advanced model yet. Beyond being smarter, it’s designed to be more reliable and versatile across a wide range of tasks. This post gives a first look at what’s new and what it means for everyday AI use.

Introducing gpt-oss

OpenAI just dropped GPT-OSS — their first fully open-source model. Open-source releases often spark waves of creativity from developers, and this could mean faster innovation (and maybe even some unexpected use cases). Worth keeping an eye on how the community runs with it.

Off The Assembly Line

One recurring challenge we noticed from real-world use: users often had project-specific details — like user manuals, FSDs, or policies — scattered across multiple documents. These details could drastically improve AI-generated test cases, but users had to share them repeatedly in every session.

To solve this, we introduced Global Instructions and Session Instructions.

• Global Instructions – Type or paste important project context (like recurring requirements, policies, or standards) once in a central place, and reuse it across the project without repeating yourself.
• Session Instructions – Add details for specific modules or features, and apply them only within a particular session of your choice.

Together, they keep context consistent and cut down on repetitive work — making test case generation faster and more relevant.