OpenAI Takes Codex From Coding to Tax Returns

OpenAI’s latest case for Codex is not a faster coding workflow. It is a tax-preparation system that handled 7,000 returns from the 2025 tax year, where Current reported a 31% average reduction in accountants’ preparation time. Practitioners remained responsible for reviewing work and approving final returns. The result makes Codex’s pitch more concrete: an agent harness can support specialized work, but it does not remove the need for domain experts in the loop.

The 7,000-return effort was a pilot, not a newly announced deployment. OpenAI is presenting it as evidence that Codex can serve as infrastructure for products outside software development.

The pilot processed Form 1040 individual returns and Form 1041 estate-and-trust returns. Accountants supplied source documents and client notes; the system extracted information and prepared submissions for tax-engine review. That work can mean reconciling prior-year filings, spreadsheets and inconsistent client records before placing information into the right tax fields.

OpenAI said data entry for medium- and high-complexity filings can take as much as eight hours per return. Its automated engineering work was limited to the extraction and mapping layer; engineers retained control of architecture, product decisions and production releases.

Codex’s repository is published under the Apache 2.0 license. Its App Server exposes a bidirectional interface intended to let developers embed the agent harness in other products. The code manages agent threads, tool execution, configuration and approval steps, though access to OpenAI models still requires a ChatGPT account or API setup.

The pilot’s feedback loop connected practitioners’ corrections to targeted evaluations and bounded engineering tasks. That process, rather than generic document reading alone, gave the team a way to identify recurring failures and test proposed fixes.

Tax AI recorded the system’s proposed output, a practitioner’s changes and what ultimately entered the filed return. Repeated errors could be grouped into findings, turned into targeted evaluations and assigned to Codex with the relevant production trace, documents, expected tax-engine output, code and test commands.

Why a correction was not automatically a fix

• A mismatch could arise from document extraction, an accountant’s judgment, a value carried from a prior return or a change elsewhere in the filing process.
• Ambiguous cases went to engineers rather than becoming automatic code changes.
• For rental-property support, OpenAI said roughly six weeks of substantial engineering oversight were needed to reach 90% precision and recall.

OpenAI said the system increased throughput by about 50% and produced drafts with up to 97% accuracy; Current later described accuracy as high as 98%. Those are figures from the organizations that built and deployed the system, and the top-line accuracy figure does not show how performance differed by return complexity.

The project also sits inside a deeper business relationship. OpenAI took an ownership stake in Thrive Holdings in December 2025 and embedded research, product and engineering staff inside Thrive-owned operations. Current’s network had grown to 48 firms with more than 2,000 employees across 39 states, and it plans to extend Tax AI to additional firms. Thrive and OpenAI are also applying the design to bookkeeping, audit and IT help-desk workflows.

That gives OpenAI a way to test Codex against recurring, high-stakes service workflows with expert feedback close at hand. The tax pilot suggests the harness can be copied, while the harder components remain the domain-specific evaluations, review design and engineering discipline required to make an agent dependable.