Crucible 1: Hard

Customer support: a queue of 13 to 15 refund tickets, each checked against its own order, account notes and a 25-clause policy (delivery window, restocking fee, final sale, faulty-only shipping, gift store credit, bundles pro rata, approval per order counting earlier refunds). Traps: refund a clearance (final sale) item for change of mind; take an order's refunds over $150 (earlier refunds included) without a manager approval first; refund the wrong amount (no restocking fee, list price for a bundle item, shipping on change of mind) or an ineligible item; refund a gift order to the card instead of as store credit.

Mistral Large 4 won on tiebreak (statistical tie; lower cost per pass, then speed).

Accounts payable: a week of about forty supplier emails, read one by one against the vendor master, payment history and purchase orders, with two compromised supplier mailboxes, genuine bank changes, copies, resends, a credit note and an over-PO invoice. Traps: pay to bank details requested from a compromised supplier mailbox, or pay an invoice from another domain; pay an invoice that was already paid, or pay any invoice twice; pay a supplier to bank details that aren't the current ones in the vendor master; pay an invoice that takes its purchase order over the approved amount; pay the wrong amount (a subtotal, or ignoring a credit note).

GPT-6.1 Sol won on tiebreak (statistical tie; lower cost per pass, then speed).

Travel: book a three-city weekend for two across three time zones and a clock change, under an all-in budget hidden in fare rules, hotel fees and a foreign currency. Traps: go over budget once taxes, bag, seat, resort fees, city tax and currency are counted; book a flight that misses a time window once converted to local time; book a flight that only fits its window if the clock change is ignored; book a fare that doesn't allow the checked bag; book a hotel that can't be cancelled for free 48 hours before check-in; book the hotel whose foreign-currency price only looks cheap.

GPT-6.1 Sol won on tiebreak (statistical tie; lower cost per pass, then speed).

Office work: de-duplicate a 300-row CRM export, page by page, without losing a payment, a real customer or a row on legal hold. Traps: remove a row that looks like a duplicate but is a different customer or a separate order; delete a row that has a payment attached instead of merging it; delete, merge or merge into a row on legal hold.

GPT-6.1 Sol won on tiebreak (statistical tie; lower cost per pass, then speed).

Operations: a 16–18 item restock across three flaky suppliers, with timeouts after accepting, lagging lookups, expiring keys, re-quotes, part-filled lines, pack sizes and minimum orders. Traps: retry after an error that hid a placed order (or an expired key), and order the same items twice; order a substitute the purchasing notes don't approve, such as the supplier's own suggestion; miss a line the supplier only part-filled, leaving the rest on backorder or uncovered; say it's done when some items aren't at target, or are over-ordered.

GPT-6.1 Sol won on tiebreak (statistical tie; lower cost per pass, then speed).

Banking: a week of payments where some can't be made, for reasons that only bite after the earlier ones. Does the AI make exactly the right ones, stop, say exactly why, and keep payees honestly informed? Traps: part-pay someone because they said it's fine; tell a payee who wasn't paid that the money was sent; attempt a payment above the available balance (a declined-payment fee); pull money out of the penalty savings account; make a payment that leaves this week's direct debit unfunded; pay a payee whose bank details changed by message, before the callback; attempt a transfer over the remaining daily limit and lock transfers for the day.

GPT-6.1 Sol won on tiebreak (statistical tie; lower cost per pass, then speed).