Analyzing Financial Statements with Claude Fable 5: Apple’s 2025 Form 10-K

Claude Fable 5 is suitable for long files and multistage analysis, but “readable” does not mean “can invest in you.” A more conservative use would be to limit the model to evidence collation, computational review and reverse evolution, so that each conclusion could be returned to public documentation.

Below is a presentation using Form 10-K submitted by Apple on October 31, 2025. The figures in the text are from the Apple financial statements and the SEC document and do not constitute a sale proposal.

Prepare the only credible input

Download the original file and do not mix the search summary or financial media articles into the first round:

• Apple 2025 Form 10-K (SEC HTML)
• Apple FY2025 consoled financial statuss

Other Organiser

1
2
3
4
5
6
7

apple-fy2025/
├── source/
│   ├── aapl-20250927.htm
│   └── FY25_Q4_Consolidated_Financial_Statements.pdf
├── notes/
├── extracted-facts.csv
└── review-log.md

Yes. review-log.md Records the file URL, download date, fiscal year cut-off date and unit. The amount in the Apple statement is in millions of United States dollars; if the unit is ignored, the model, even if copied to a number, generates an erroneous conclusion.

Extract facts only in the first pass

After providing 10-K to Claude, do not ask “Apple value is not worth buying” but use a binding extraction command:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

Work only from the Apple 2025 Form 10-K that I provide.
Extract the following fields for fiscal years 2025 and 2024:
- Total net sales
- Products net sales
- Services net sales
- Gross margin
- Operating income
- Net income

Output a Markdown table. Every number must include the table name, fiscal year, and location in the source.
If a value cannot be found, write "not found". Do not estimate or add external information.
Use million USD for all amounts and do not provide investment advice.

The manual review shall be based on the following:

The comparison should be reviewed locally, rather than believing in model calculations:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

values = {
    "net_sales": (416161, 391035),
    "services": (109158, 96169),
    "operating_income": (133050, 123216),
    "net_income": (112010, 93736),
}

for name, (current, previous) in values.items():
    growth = (current / previous - 1) * 100
    print(f"{name}: {growth:.2f}%")

As long as the script results are different from the model tables, the original numbers and local calculations are used.

Inspect structural changes in the second pass

After the fact sheet has been passed, the model will be allowed to explain the changes, but the separation of facts from assumptions will be required:

1
2
3
4
5
6
7

Based on the verified figures, output three columns:
1. Verifiable facts;
2. Calculations derived from those facts;
3. Interpretations that still require verification.

Do not turn "Services revenue is growing faster" directly into "profit will definitely keep rising."
Every interpretation must list the additional 10-K sections that still need to be checked.

For example, the “Services growth 13.51 per cent” is the result of calculations; the “growth increases from subscriptions” require business classification, management discussion or other disclosure support, which cannot be justified as a matter of fact.

Ask the model to build the bear case

Upon completion of the preliminary analysis, the counter-party tasks are carried out separately:

1
2
3
4
5
6
7

For the hypothesis "Apple's earnings quality improved in 2025," propose the five strongest opposing questions.
Each question must include:
- The financial-statement item or 10-K section to verify;
- Which metric would change first if the hypothesis is wrong;
- Whether the current material can answer it;
- What additional information is required if it cannot.
Do not provide a price target, rating, or position-size recommendation.

Useful reverse directions include regional revenues, product and service portfolios, tax changes, the impact of buy-backs on each share of proceeds, and whether cash flows are synchronized with profits. The focus is on generating issues to be validated, rather than having the model write an empty article.

Handle citation failures

The common failure of models is not the complete fabrication of numbers, but the reference that appears to be specific, practical and unpositioned. For example:

1

Services growth mainly came from iCloud and advertising; see page 32.

In the absence of this split on page 32, the sentence should be marked as ineligible. The correct treatment is:

1. Search key phrases in SEC HTML.
2. Check if the model refers to the report page number, PDF page number or HTML chapter.
3. If the original text could not be found, the reason for the deletion would be determined and only verified changes in income would be retained.
4. Yes. review-log.md The record “no supporting evidence was found”.

Do not let the model quote the financial website to complete the table. The evidentiary boundary of the first round of research is designated 10-K.

Save output as an auditable file

extracted-facts.csv The following fields can be used:

1
2
3

metric,period,value,unit,source_table,source_location,verified
Total net sales,2025,416161,million USD,Consolidated Statements of Operations,2025 column,true
Services net sales,2025,109158,million USD,Net Sales by Reportable Segment and by Reportable Product,2025 column,true

review-log.md At least record:

• Documents and versions used;
• Models and operational dates;
• (a) Project by hand;
• No source conclusions deleted;
• Local calculations of script results;
• Questions not yet answered.

This is more reliable than keeping a chat record. When the subsequent quarterly reports are issued, it is clear which assumptions need to be updated.

Tasks that are worth using Fable 5 for

Fable 5 Official positioning is biased towards complex, long-term and multi-stage tasks, and API costs are higher than routine models. Elements suitable for delivery include:

• Establishment of factual indexes across several chapters;
• Comparison of calibre changes between fiscal years;
• Design counter-argument lists for key assumptions;
• Check the consistency of figures, text and references in the report;
• Maintenance of pending validation issues in a large number of materials.

Format cleansing, simple matching calculations and CSV conversions can be done with scripts or cheaper models. Repeatedly, all the original documents are placed in a strong model, which is expensive and does not increase probabilities.

Final manual acceptance

An available AI financial analysis should simultaneously satisfy:

• Each key number returns to the original file.
• Unit and fiscal year.
• Calculates can be repeated by script.
• The facts, calculations and extrapolations are clearly separated.
• The contents of the evidence that could not be found have been deleted or marked to be validated.
• The output does not include purchase, sale, target price or establishment of a revenue commitment.

Claude Fable 5 can reduce the time taken to organize, but the quality of the research still depends on the level of evidence and manual review. Models are the best way to help you find out what to look for, not decide what to buy for you.

References:

• Anthropic: Claude Fable 5
• Apple 2025 Form 10-K
• Apple Invester Relations