Claude Fable 5.2 can write more complex SQL. What should it calculate in a sales report?

Public developments and original analysis · Announcement date: 2026-11-09 · Sources

Anthropic released Claude Fable 5.2 on November 9. Its announcement described a trial by an Oracle machine learning lead involving country debt summaries, repayment histories and regional benchmarks. According to the account, Fable 5.2 returned the expected result. SQL from two other models also ran successfully, but counted country-level totals more than once. The announcement describes one test and provides too little detail to reproduce it independently. It illustrates a concrete problem for reporting teams. The database can return a number without knowing what the team intended to count.

Fable 5.2 is available through the Claude API, and AWS confirmed availability on its platform the same day. Anthropic lists standard pricing at $8 per million input tokens and $40 per million output tokens. For a reporting team, the decision also depends on how long the old report took to verify and how much of that work a model change would save. Recalculating every figure that finance or analytics already approves means reopening settled questions every week.

Sales by order or by line item

Consider a fictional retail report. The sales lead asks for last week’s net sales by channel and the sales attributable to a new product. The orders table has one row per order. The items table has one row per line item. Suppose an order for 300 contains 200 of the new product and 100 of an older product. Once the two tables are joined, both item rows carry the same order total of 300. Sum the order total across both rows and the result is 600. Filter to the new product and sum the order total, and the result is still 300. That answers how much customers spent on orders containing the new product. In this simplified example, sales of the new product itself are 200. Both questions can be useful, but confusing them can change the decision about the launch budget.

Refund dates can change the meaning of net sales, too. For an order sold last week and refunded this week, the team can deduct the refund in the week it occurred or restate the original sales week. The first view helps track what changed this week. The second helps measure what remained from a specific cohort of sales. If finance uses the first definition and marketing uses the second to judge a launch, label each number accordingly. Asking a model to guess the rule behind an old spreadsheet can hide the disagreement in a query that reviewers rarely see.

Separate the weekly report from one-off questions

For recurring metrics, start with an approved data view, query or existing report. Pass the metric definition, refund treatment and update time along with the figures. The model can then write the commentary and surface questions worth investigating. If net sales fell but the order count held steady, it can help examine average order value, refunds and the channel mix separately. The report should point readers to the relevant figures so they can follow up. If the team simply needs a fixed table every week, a subscription to the existing BI report may be all it needs.

When the figures come from different systems, a late feed can look like a business change. On Monday morning, online sales may be current while Sunday’s store transactions are still syncing. If this week and last week use different cutoffs, the model may interpret missing sales as a decline. Show when each channel was last updated and mark incomplete figures separately. Wait for the missing data or compare only periods that are complete on both sides.

An ad hoc question is a better place to try Fable 5.2. A channel manager might ask which customers bought from the company for the first time during a campaign when there is no saved query for that question. Ask the model to specify whether it plans to count orders, customers or items, and to propose the fields and exclusions. An analyst can compare the result against historical samples in a read-only environment. If it is unclear how to identify the same customer across channels, resolve that definition first. When the question recurs and the query has passed review, consider saving it for the team to reuse.

For businesses that need figures from multiple systems, Mokyun AI’s integration offering provides a basis for assessing how to retrieve approved order, refund and channel results through existing interfaces, files or authorized data views. The data team defines the metrics. A digital employee can assemble the report, describe changes and track discrepancies that still need an answer. Verify the data sources, model channel and applicable permissions in the deployment environment. De-identified historical data can show whether this division of work saves time before live sales data is connected.

An initial comparison can be small. Use examples with multiple items, refunds in a later week and missing channel data to check that the fixed report still matches approved figures. Then give the incumbent tool and the new model the same ad hoc questions. Record the share of acceptable answers, the analyst’s review time and the cost of the calls. If the errors come from undefined metrics, fix the report first. Give a stronger model a dedicated place once the definitions are settled and difficult follow-up questions still consume substantial analyst time.

• Mokyun AI systems integration

Mokyun Editorial · neo@mokyun.com