How Finance Analysts Use Claude Fable 5 to Extract 10-K Tables

Table of Contents

I still remember the sound. My friend Sarah, a senior financial analyst at a mid-size hedge fund in New York, slumped her head onto her keyboard at 11:47 PM. Thud. The hollow echo of exhaustion.

She’d spent the last nine hours manually extracting data from a 400-page annual report. Not reading it. Just copying numbers from 37 different tables, cross-referencing footnotes, and reformatting everything into Excel. Her firm paid her $140,000 a year to find investment insights. Instead, she was doing digital data entry.

That’s the dirty secret of finance. Analysts don’t spend most of their time analyzing. They spend 60-70% of their week cleaning, extracting, and wrestling with PDFs, scanned tables, and inconsistent chart formats.

Then I showed her Claude Fable 5.

We ran a real test together last Tuesday. She handed me a messy 10-K filing from a struggling retail chain. I fed it into Fable 5. Eleven minutes later, we had a clean, structured dataset, a summary of three key margin anomalies, and a ready-to-paste table for her pitch deck.

She stared at the screen. Then she looked at the clock. Then she started laughing.

That’s the moment I knew I had to write this down for every other analyst stuck in the same grind.

The Executive Workflow Summary (Read This First)

• Target Persona: Financial Analyst (buy-side or sell-side), Equity Researcher, Investment Associate.
• The Old Bottleneck: 6-9 hours per 10-K/10-Q manually extracting tables, fixing OCR errors, and aligning data across pages.
• The New AI Workflow: Claude Fable 5 (1M token context + native table/chart vision) ingests the full PDF, extracts all structured data, and outputs clean JSON/CSV with cited cell locations.
• The Measurable ROI: 11 minutes of AI oversight vs. 7.5 hours of manual grunt work. $190 saved per report in analyst time (assuming $120/hr loaded cost). That’s $9,880 per quarter for a typical coverage universe of 52 companies.

Why I Got Obsessed With This Problem

I’m not a finance guy. I’m a tech writer who hangs out with too many finance people. But three weeks ago, my cousin Mike (portfolio analyst at a NYC asset manager) called me at 9 PM sounding like a ghost.

“Dude,” he said. “I just spent my entire weekend copying numbers from 14 different ESG reports. My wife is going to divorce me. There has to be a better way.”

That stuck with me. I started asking around. Every single analyst I spoke to—from sell-side juniors to hedge fund quants—complained about the same thing. The data is locked inside PDFs. Tables get split across pages. Charts are just images. And the “smart” tools like Tabula or Camelot still break when a table has merged cells or weird spacing.

So I made a deal with Mike and Sarah. I’d pay for the API credits. They’d give me a real, messy, work-in-progress PDF that had been making them miserable. We’d test Claude Fable 5 together, live, in my apartment.

No cherry-picking. No clean examples. Just the ugly reality of financial research.

Phase 1: The Real Bottleneck (It’s Not What You Think)

The obvious answer is “PDFs are hard to parse.” But that’s too simple.

Here’s what I learned from watching Sarah work on her laptop for three hours.

The real time-suck isn’t the extraction itself. It’s the validation loop. You copy a number from page 27, table 4. Then you check page 28, footnote 12, because the table says “see note 12 for adjustment.” Then you realize the note says “excluding one-time legal fees,” so you have to manually subtract that from the original cell. Then you do that 400 times.

By hour five, your eyes glaze over. You start making typos. Then you have to go back and reconcile.

Mike put it perfectly: “I don’t hate the work. I hate the second-guessing. Did I grab the right column? Did the OCR turn ‘85,000’ into ‘88,000’ because that scanned ‘5’ looks like an ‘8’?”

That’s the hidden tax. The mental cost of distrusting your own data entry.

And the tools they currently use? Adobe Export PDF loses formatting. Python libraries like camelot-py work 70% of the time but crash on nested headers. Abbyy FineReader costs $600 and still requires hand-holding.

So analysts default to manual. Because manual is reliable. It’s just soul-crushing.

Phase 2: Why I Picked Claude Fable 5 Over Every Other AI

I tested three alternatives before settling on Fable 5.

• GPT-5.5 Pro: Good at understanding tables, but the 128k token limit meant I had to chunk PDFs. Chunking breaks cross-page references. Pass.
• Gemini 3.1 Pro: Handles 2M tokens, but its table parsing is sloppy. It invents numbers when a cell is empty. Dangerous for finance.
• Claude 3.7 Sonnet: Solid, but it struggles with rotated text or tables that span columns.

Fable 5 won for three specific reasons that matter to analysts:

• The 1M context window: It swallowed an entire 500-page 10-K in one go. No chunking, no lost cross-references.
• Native vision + text fusion: When a table is actually a scanned image (common in older filings), Fable 5 reads the pixels and the surrounding text simultaneously. It caught a footnote that said “see page 142 for restated figures” and automatically adjusted the numbers.
• Structured output without hand-holding: I didn’t have to write a complex prompt. “Extract all tables from this PDF as JSON, include page numbers and footnote references” worked on the first try.

Sarah was skeptical. “AIs hallucinate,” she said. “I can’t put fake numbers in a model.”

I told her we’d verify everything. That’s Phase 4.

Phase 3: The Live Test (A Real 10-K From Hell)

We used a 2025 annual report from a mid-cap apparel retailer. I’d rather not name the company, but trust me—the filing was a disaster.

• 487: pages.
• 63: separate tables.
• 12 tables: were scanned images (no selectable text).
• 5 tables: had merged cells and multi-line headers.
• Footnotes: scattered across 30 different pages.

The workflow we executed:

• I downloaded: the PDF from the SEC EDGAR system (free, public).
• Uploaded it directly: to Claude Fable 5 via the API (not the web chat—the API allows file uploads up to 100MB).
• Used this exact prompt:

“You are a financial data extraction expert. This PDF is a 10-K. Extract every single table, including footnotes. For each table, output: the page number, the raw data as a CSV array, and any footnote text that modifies specific cells. Also flag any numbers that have ‘restated’ or ‘adjusted’ next to them. Return everything as a single JSON object.”

Hit enter. Went to make coffee. Came back 9 minutes later.

The final result: A 1.4MB JSON file containing all 63 tables, fully structured. Every footnote was linked to the relevant cell coordinates.

We spot-checked 15 random cells across different tables. Every single number matched the original PDF. The one “error” we found wasn’t an error—the PDF itself had a typo (listed “$4.5 million” as “$4.5 millon”), and Fable 5 preserved the typo exactly as shown.

That’s actually perfect. You want the AI to transcribe, not correct.

Total time from PDF upload to usable JSON: 11 minutes and 22 seconds.

Sarah’s normal manual time for this filing: 7.5 hours.

Phase 4: Where the Magic Stops (You Still Need Human Eyes)

I’m not going to pretend Fable 5 is perfect. It’s not.

We found two categories of problems during our test.

The first: footnote ambiguity. One table had a footnote that said “*excluding store closures in Q3.” The footnote was attached to a single number (revenue). But the same footnote symbol appeared next to a different number on the next page. Fable 5 correctly flagged the conflict but couldn’t resolve it. Sarah had to read the actual text to understand that the footnote applied to both numbers.

The second: chart data. Fable 5 is bad at extracting precise numbers from bar charts or line graphs. It gave us approximate values (“looks like $12.3 million”) when the chart was low-resolution. We had to skip those and pull from the table version instead.

What Sarah had to recheck manually:

• Every “restated” footnote: (6 of them) to ensure the adjustment logic was applied correctly.
• Any table: where the column header was split across two rows (happened twice).
• All chart-based data: (we just ignored it and used the table supplements).

That took her about 45 minutes. Still a fraction of the 7.5 hours she would have spent copying everything from scratch.

The rule we settled on: AI does the heavy lifting. Human does the cross-check on footnotes and structural oddities.

Phase 5: What Sarah Decided After the Test

I asked her point-blank: “Are you switching?”

She didn’t hesitate. “Yes. But not for everything.”

Here’s her new workflow starting this week:

• Use Fable 5 for: Any 10-K or 10-Q longer than 100 pages. Any filing with more than 20 tables. Any PDF that’s a scan (no selectable text).
• Still do manually for: Short documents (under 20 pages). Quarterly reports where she only needs one or two specific numbers. Anything with complex legal footnotes that require interpretation.

She also bought a $50 API credit pack (her firm reimbursed it). She estimates she’ll use about $12 worth per week, covering 8-10 filings.

That’s $624 per year in API costs. Compared to the 300+ hours she’ll save, it’s a no-brainer.

The Workflow ROI Comparison Table

Workflow Stage	The Manual Way (Time)	The Fable 5 Way (Time)
Download & open PDF	1 min	1 min (same)
Scan for tables & charts	15 min (skimming)	0 min (AI auto-detects)
Extract first table	8 min (copy, paste, clean)	0 min (batch extraction)
Extract 63 tables (batch)	340 min (spread across 5-6 hours)	9 min (AI processing)
Handle scanned/image tables	45 min (retype or OCR + proof)	0 min (AI vision)
Cross-check footnotes	30 min (manual lookup)	30 min (still human)
Validate 15% random cells	20 min (spot-check)	10 min (faster with AI output)
Format into Excel/CSV	15 min	2 min (AI already structured)
TOTAL	474 min (7.9 hours)	52 min (0.87 hours)

Time saved per 10-K: 7.03 hours.

The Pricing Reality: Subscription vs. Analyst Hourly Rate

Claude Fable 5 API: $10 per 1M input tokens / $50 per 1M output tokens.

A typical 500-page 10-K is about 180,000 input tokens (including the PDF text and table metadata). Output (JSON with 63 tables) is roughly 90,000 tokens.

Cost per filing:

• Input: 180k tokens × $0.00001 = $1.80
• Output: 90k tokens × $0.00005 = $4.50
• Total: $6.30 per 10-K

Compare that to an analyst’s fully loaded hourly cost ($120/hr including benefits, software, desk costs).

• Manual: 7.9 hours = $948 of labor.
• AI-assisted: 0.87 hours = $104 of labor + $6.30 API = $110.30.
• Savings per filing: $837.70.

If an analyst covers 150 filings per year (reasonable for a sector specialist), that’s $125,655 in annual labor savings for one person.

Before vs. After: The Stress & Difficulty Reality Check

Task	Manual Method (Stress 1-10)	Using Claude Fable 5 (Stress 1-10)
Extracting tables from scanned PDFs	9 (constant fear of OCR errors)	3 (AI handles it, but I still spot-check)
Reconciling footnotes across pages	8 (tedious, easy to miss)	5 (AI flags them, I just verify)
Maintaining data sanity (no typos)	7 (one typo ruins a model)	2 (AI output is consistent)
Meeting tight deadlines before earnings	9 (panic mode every quarter)	4 (I can run multiple filings in parallel)
Mental fatigue after 6 hours of PDF work	10 (braindead)	3 (I do 30 minutes of checking then move on)

Sarah’s comment after seeing this table: “The stress drop from 9 to 3 on scanned tables alone is worth the API cost. I used to dread filings with bad OCR. Now I almost look forward to them.”

The Adoption Scalability Verdict (My Honest Rating)

How easy is this for a typical finance team to implement permanently?

I’d say 8/10. The only friction points are: getting API access (Anthropic takes 1-2 days to approve), training analysts to write good extraction prompts, and building a simple script to call the API from Excel or Python. A half-decent quant could set this up in an afternoon.

The disadvantages of using AI (and how Sarah overcomes them):

• Footnotes: still need human interpretation. She overcomes this by always reading the footnote section manually after the AI extracts. Takes 20-30 minutes. Non-negotiable.
• Chart data: is unreliable. She simply doesn’t use Fable 5 for charts. She pulls those numbers from the table supplements instead.
• Occasional hallucination: on empty cells. She always spot-checks 5-10 random cells per filing. Takes 10 minutes.

The disadvantages of the old manual method:

Everything. It’s slow, error-prone, and burns out analysts. Sarah said she would never voluntarily go back to manual extraction for long documents. The only time she uses manual now is for a 5-page press release where opening Fable 5 takes longer than just typing the three numbers she needs.

My personal rating of Claude Fable 5 for this use case:

9/10. Deducting one point only because of the chart limitation and the need to verify footnotes. For pure table extraction from messy PDFs? It’s the best tool I’ve ever seen.

FAQ (What Every Finance Pro Asks Before Ditching Manual Work)

Can Claude Fable 5 handle password-protected or encrypted PDFs from sell-side research?

No. You must decrypt the PDF first (standard Adobe password removal). Fable 5 has no built-in decryption for security reasons. For confidential internal documents, Anthropic’s API is GDPR and SOC 2 compliant, but check your firm’s data policy before uploading.

Does Fable 5 work with scanned handwritten tables (e.g., old annual reports from the 1990s)?

Partially. Clear handwriting (think typed manuscript) works about 80% of the time. Cursive or faded scans fail badly. For those, you’re still better off with a human data entry contractor.

How do I know Fable 5 isn’t hallucinating numbers?

You don’t. That’s why you spot-check. The API can return “confidence scores” for each extracted cell when you ask. I recommend adding “include a confidence score (0-100) for every numeric value based on visual clarity and context” to your prompt. Anything under 85% gets flagged for manual review.

What’s the maximum PDF file size Fable 5 can handle?

Via API, 100MB. Via the Claude web interface, 30MB. A typical 10-K is 5-15MB. Very large filings with hundreds of images might hit the limit—then you need to split the PDF into two parts.

Can I use this for real-time earnings call transcripts (not just 10-Ks)?

Yes, and it’s even faster because transcripts are mostly clean text. Fable 5 can extract guidance tables and Q&A numbers in about 3 minutes. I’ve tested it on four transcripts. Works perfectly.

Is there a risk of Anthropic training on my financial data?

By default, no. API data is not used for training unless you opt in. But read the fine print. Enterprise customers can sign a data privacy addendum. I recommend doing that before uploading any non-public information.

The Annual Savings Math (The Only Number That Matters)

Let’s do the final calculation for a typical analyst covering 150 filings per year.

Manual method:

• 7.9 hours: per filing × 150 filings = 1,185 hours
• 1,185 hours: × $120/hour (loaded labor) = $142,200 per year

Claude Fable 5 method:

• 0.87 hours: per filing × 150 filings = 130.5 hours
• 130.5 hours: × $120/hour = $15,660 labor
• Plus API costs: 150 filings × $6.30 = $945
• Total: = $16,605

Annual savings: $142,200 – $16,605 = $125,595

That’s not a “nice to have.” That’s a 46% increase in effective analyst capacity without hiring another person.

Or put another way: every analyst who adopts this workflow frees up 1,054 hours per year to do actual analysis—building models, talking to management, finding mispriced assets. The work that actually makes money.

Sarah’s final words to me after I ran these numbers: “I’m sending this to my head of research tomorrow. If we don’t do this, we’re leaving money on the table.”

That’s the punchline. Fable 5 isn’t just a tool. It’s a competitive advantage. And right now, most finance teams haven’t even heard of it.

Don’t be most teams.