Modeling complex taxes with Claude Code (1M context)

Fed a full year of tax documents — K-1s, dozens of 1099s, bank and brokerage statements, prior-year returns, a spouse's W-2, medical records, business expenses and mortgage statements — into Claude Code with the 1M-token context model, asking it to estimate the state and federal tax bills an accountant would calculate (new SALT limits included). A regular Opus 4.6 chat with extended thinking couldn't handle it and ChatGPT Pro gave up after 40 minutes; Claude Code plowed through in under five minutes, landing within 0.5% of the accountant's final numbers, and produced a detailed spreadsheet with every assumption plus what-if scenarios to test.