OpenAI reports Basis’s faster tax-workbook test with GPT-6 Astra

OpenAI reports Basis’s faster tax-workbook test with GPT-6 Astra

OpenAI’s page, dated September 28, 2026, reports Basis’s comparison of GPT-6 Astra and GPT-5.6 Sol on a complicated tax workbook with 50 tabs. Basis says GPT-6 Astra completed it in half the time in its test. Basis also reports about a 20% improvement in internal evaluation scores, assessing agents’ work and final answers, including template-following, use of primary sources for tax questions, and checking their work. Basis attributes the gains to better understanding of user intent, including when to ask questions and flag assumptions, and says the model can infer expectations from broader context with fewer explicit instructions. The page says this gives Basis more confidence its agents can handle situations beyond internal tests; it does not report results from those situations. Basis also says adjusting reasoning effort during a task helps reduce cost and response time, but provides no measurements for those effects. The evidence describes a potential change in accountants’ work, but does not report changes to jobs or time spent on other tasks.

Publication: cb2a832d-d4aa-11a8-8402-417ae71edc0f

attribution — pass: This is a single OpenAI customer case study about Basis. The reading attributes claims to OpenAI's page, to Basis, and implicitly to co-founder Troyanovsky's statements. It does not present them as established facts.

“GPT-6 Astra is able to complete that workbook in half the time that GPT-5.6 Sol is able to.”

summary — uncertain: The reading never says who Basis is: a startup that builds AI agents to automate accountants' manual work. Without that, a general reader cannot see why a tax workbook test matters, or why the closing sentence turns to accountants' work. 'Agents' is also central and unexplained, and no definitions are supplied. Smallest repair: add one sentence identifying Basis and briefly explaining what an AI agent is here. The reading's treatment of the product stays with claimed effects and remains attributed, so it does not read as promotion.

Basis⁠(opens in a new window) builds AI agents to automate much of the manual work that accountants do each day

date — pass: The page dateline, September 28, 2026, matches originalPublished, which is stated at day precision. The archive copy is disclosed in the limitations. The 'Keep reading' items dated Sep 29, 2026 are probably later page additions; the reading neither uses nor relies on them. There is no hindsight framing.

September 28, 2026 Basis completes a tax workbook 2x faster with GPT‑6 Astra

support — pass: Each claim matches the source: the 50-tab workbook finished in half the time, the roughly 20% gain in internal evaluation scores, the evaluation criteria, the cost and latency claim for adjusting reasoning effort, and the stated confidence beyond internal tests. The caveats about missing methodology, measurements and real-world results are accurate. The cautious closing about accountants' jobs adds nothing unsupported.

Basis saw about a 20% improvement in its internal evaluation scores with GPT‑6 Astra, driven by better understanding of user intent

media — pass: No media candidates were supplied and none was selected, which is appropriate.