We read Anthropic’s S‑1 the way Wall Street will

Anthropic, PBC filed its registration statement with the SEC this morning: 420 pages of prospectus and 34 exhibits. Before most people finished the cover page, bem had classified it, cut it into sections, pulled out every number an analyst needs and checked its own work. It also matched every counterparty against public comps and SEC EDGAR, built a knowledge graph across the exhibits, drafted an initiation-of-coverage note as a Word document, and stood up a chatbot that answers questions with page citations.

This post walks through every step, with the real outputs. Everything below was produced by bem’s public v3 API in production, on the filing as it appeared on EDGAR.

The document: 420 pages, 34 exhibits, and a lot of blanks

An S-1 is the hardest document in finance to read quickly. It mixes legal prose, risk disclosures, audited tables, marketing gatefolds, charts drawn as images, and dozens of exhibits, each a contract in its own right. Anthropic’s is no exception. Many of its biggest numbers are left blank (“$ ”) until the offering is priced.

We downloaded the filing (accession 0001628280-26-067168) and rendered it to a 30 MB PDF that keeps EDGAR’s page layout. Below is every page, colored by how bem’s split function divided it: no page numbers, no table of contents, only the content. Hover a page to see its section.

One workflow, ten function types

The analyst desk is a single bem workflow, anthropic-s1-analyst, with 22 nodes. One PDF goes in. It fans out across eight sections, runs nine extractors in parallel, enriches the outputs, grades them, reshapes them, and delivers them. Two companion workflows handle the rest: anthropic-s1-memory parses the S-1 and exhibits into a navigable knowledge graph, and anthropic-s1-initiation-note joins every extraction into a research note and renders it to .docx.

Classify, then split where a human would

Classify looks at the cover and routes the file. A real desk receives 10-Ks, 10-Qs, 8-Ks and proxies through the same entry point. This one came back s1_registration_statement in 29 seconds, so the workflow sent it down the S-1 branch.

Split (semantic_page) then read all 420 pages and cut them into 20 contiguous pieces across nine section types, in 2 min 47 s. Each piece went to its own extractor. Every boundary matches the printed table of contents: Risk Factors runs from printed page 26 to 120 (PDF pp. 37–131), and the F-pages start at PDF p. 346.

{
  "functionName": "s1-section-splitter",
  "type": "split",
  "splitType": "semantic_page",
  "semanticPageSplitConfig": { "itemClasses": [
    { "name": "risk_factors", "description": "The 'RISK FACTORS' section: bolded risk headings and their explanations", "nextFunctionName": "s1-risk-factors" },
    { "name": "mdna", "description": "MD&A: key metrics, results of operations, liquidity, contractual obligations", "nextFunctionName": "s1-mdna-kpis" },
    { "name": "financial_statements", "description": "Audited financial statements: pages numbered F-1 onward", "nextFunctionName": "s1-financial-statements" }
  ]}
}

Extract: every number, with a pointer to where it came from

Nine extract functions ran in parallel, each with its own focused JSON Schema: offering terms, financial statements, MD&A KPIs, compute commitments, risk factors, business counterparties, governance and pay, ownership, and offering mechanics. MD&A feeds two of them, because one schema that tries to do everything does each job worse. The financial statements run with enableBoundingBoxes, so every value carries its location on the page.

Try it: click any number below, and bem’s bounding box lights up on the actual page of the audited income statement.

What the numbers say

H1 2026 shows a $80.7B net loss on $16.3B of revenue. But 90% of that loss is the remeasurement of convertible notes and SAFEs, a figure that grows as Anthropic’s valuation rises. The operating loss was just $693M, and operating cash flow turned positive at +$3.9B, against −$1.5B a year earlier.

Data that only exists inside a picture

The MD&A’s best numbers aren’t in the text at all. The quarterly revenue, compute spend and margin series exist only as charts embedded as images. bem read them off the images. Below, on the left, is the original chart from the filing; on the right are the series bem extracted, re-plotted.

The compute commitments

A dedicated extractor pulls every named compute, lease and hardware arrangement out of MD&A. The prospectus leaves the aggregate blank pending pricing, but the individual deals are all there.

Enrich: risks to a taxonomy, counterparties to tickers

A list of risk factors doesn’t make an analysis. To compare companies you have to map each risk onto a fixed taxonomy. We loaded a 22-category sell-side risk taxonomy into a bem collection. Each category carries an analyst severity weight and the question we’d ask on the roadshow. An enrich function then matched every extracted risk to it with hybrid search and an LLM re-rank.

A second enrich function took every company the Business section names and ran two lookups. It searched a public-comps collection, then called SEC EDGAR’s company search live as an external endpoint, with an LLM choosing the right registrant. Matches the re-ranker didn’t confirm are left blank. That’s the right answer for private firms, and for names too ambiguous to resolve.

Evaluate: a second model grades every field

We wired evaluation nodes behind the financial statements, MD&A, compute commitments, offering snapshot and risk factors, and behind the join that writes the note. Each one grades every field: a confidence score, a hallucination flag, a written rationale, and a suggested value when it disagrees. Evaluation also runs on demand (POST /v3/eval) against any past extraction.

The judge earned its keep, and so did iterating. On the first run it flagged that the offering snapshot had captured only 6 of the 22 underwriters in the syndicate, and that the MD&A revenue mix was incomplete. Those flags fed straight into new versions of the functions: every PATCH /v3/functions/{name} creates an immutable version, and PATCH /v3/workflows/{name} pins the workflow to them. We also learned to keep each schema focused. Pulling compute commitments into a dedicated extractor took coverage from six of the compute deals to all nine.

Shape, send, join, render: from fields to a finished note

Payload shaping runs JMESPath with no model in the loop. It turns the financial statements into chart-ready series and the governance extraction into graph edges, and send nodes deliver both downstream as signed webhooks.

Then the second workflow takes over. A join fuses the nine extractions into one schema: thesis, bull and bear cases, top risks, governance, compute, and roadshow questions. Payload shaping converts that into render primitives, and render fills a Word template we designed. The result is a real .docx initiation-of-coverage note, delivered by send, with its own LLM judge on the join.

// s1-initiation-note-shape: JMESPath, no LLM
{ TITLE: title, RATING: rating,
  THESIS:      thesis[*].{paragraph: {text: @, styleId: 'Normal'}},
  KEY_NUMBERS: [{table: {rows: keyNumbers, styleId: 'TableGrid'}}],
  BULL_CASE:   bullCase[*].{paragraph: {text: @, styleId: 'ListBullet'}},
  ROADSHOW_QUESTIONS: roadshowQuestions[*].{paragraph: {text: @, styleId: 'ListNumber'}} }

Download the rendered note (.docx). Every number in it comes from the extractions. Where the S-1 leaves a figure blank pending pricing, the note says so.

Parse, ontology, knowledge graph: who is connected to whom

Extraction answers questions you already know to ask. Parse is for the ones you don’t. It turns a document into page-aware sections plus named entities and the relationships between them, then links entities across documents into one memory. We parsed the prospectus and 11 exhibits into a dedicated bucket. The exhibits include the LTBT trust agreement, the Founders LLC agreement, the Broadcom equipment lease, the Morgan Stanley credit agreement, the convertible note facilities and the stockholders’ agreement.

Before parsing we defined an ontology: 22 entity types with a hierarchy and attribute schemas. We then seeded 51 canonical entities with synonyms, so “LTBT”, “the Trust” and “Anthropic Long-Term Benefit Trust” resolve to one node from the first page.

• 12 documents parsed
• 1,035 pages
• 6,644 page-aware sections
• 2,225 entities (per-doc)
• 1,686 relationships with evidence

A chatbot on the File System API

Parsed documents are addressable through bem’s File System API (POST /v3/fs), using Unix verbs an LLM already knows: ls, grep, cat, head, stat, and the memory ops find, open and xref. Give Claude a single tool that wraps it and you have an analyst’s assistant that greps for a topic, reads the pages, follows an entity into an exhibit, and cites what it read. There’s no chunker, no embeddings and no retrieval tuning.

def bem_fs(args: dict) -> str:
    body = {**args, "context": {"bucket": BUCKET}}
    r = requests.post(BEM_FS, headers={"x-api-key": os.environ["BEM_API_KEY"]}, json=body, timeout=120)
    text = r.text if r.ok else f"HTTP {r.status_code}: {r.text}"
    if len(text) > MAX_RESULT_CHARS:  # keep tool results small; Claude can page with range/limit
        text = text[:MAX_RESULT_CHARS] + "\n…[truncated — narrow with range.pageRange, path, or limit]"
    return text


def ask(question: str, client: anthropic.Anthropic, trace=print) -> str:
    messages = [{"role": "user", "content": question}]
    while True:
        response = client.beta.messages.create(
            model="claude-opus-5-5",
            max_tokens=16000,
            thinking={"type": "adaptive"},
            output_config={"effort": "high"},
            betas=["server-side-fallback-2026-07-01"],
            fallbacks="default",
            system=[{"type": "text", "text": SYSTEM, "cache_control": {"type": "ephemeral"}}],
            tools=TOOLS,
            messages=messages,
        )
        if response.stop_reason == "refusal":
            return "(declined)"
        messages.append({"role": "assistant", "content": response.content})
        if response.stop_reason != "tool_use":
            return "".join(b.text for b in response.content if b.type == "text")
        results = []
        for block in response.content:
            if block.type == "tool_use":
                trace(f"  → bem_fs {json.dumps(block.input)}")
                results.append({"type": "tool_result", "tool_use_id": block.id, "content": bem_fs(block.input)})
        messages.append({"role": "user", "content": results})

Below are six real, unedited sessions. Pick a question to see the File System API calls the model made, then its answer.

Images, too: twenty customer one-pagers

The Business section includes 20 customer case studies, each a designed image. A separate anthropic-s1-case-studies workflow sent each JPEG through one extract function, pulling out the customer, the products, every quantified result and the quote.

Fun facts

• 1 share = 50.1%. Before the offering closes, Anthropic issues exactly one share of Class F common stock, to an LLC owned by its seven co-founders. That single share carries 50.1% of the voting power on ordinary matters. It has no economic rights. (Cover page · PDF p.2)
• Five classes. Class A (one vote), Class N (non-voting, for strategic partners), Class T (held by the Long-Term Benefit Trust, which elects a board majority), Class F (the founders’ share) and Class E (employees, limited votes). Only A and N carry economic rights. (Cover page · Description of Capital Stock)
• $417B. MD&A leaves the aggregate commitment blank pending amendment. A subsequent-events note in the audited financials states it anyway: about $417 billion of new future minimum payment obligations since December 31, 2025, with terms of up to 20 years. (Notes to financial statements · subsequent events)
• $161.2B. Anthropic’s TPU equipment leases with Broadcom total about $161.2B of rent excluding interest. That is larger than any of its cloud deals: Google about $114.5B, Amazon $110.0B, Microsoft $44.9B. (MD&A · Infrastructure Service Arrangements)
• “Render unusable”. If Anthropic defaults under the Broadcom master lease, Broadcom may take possession of, or “render unusable”, any pod of equipment, “without any court order”. (Exhibit 10.11 · Section 14.02)
• 0.40× → 2.08×. Revenue per dollar of compute rose more than fivefold in five quarters, from Q1’25 to Q2’26. Adjusted operating margin went from −221% to +19%. These numbers appear only in chart images, which bem read. (MD&A charts (images))
• ~900%. Net revenue retention as of September 30, 2026. Customers who spent $100K+ grew from about 20 at the end of 2023 to more than 11,900. (Prospectus summary · MD&A)
• 90%. of the $80.7B H1 2026 net loss is the change in fair value of convertible notes and SAFEs, which grows as the valuation rises. Operating cash flow was +$3.9B. (Statements of operations · PDF p.352)
• 271 vs 22. Firefox vulnerabilities found by Claude Mythos Preview, against 22 found by Claude Opus 4.6. On ExploitBench, scores went from 40% (Opus 4.8) to 91% (Opus 5.5) in about four months. (Mozilla case study (image) · Business)
• v. Department of War. In February 2026 the Secretary of War designated Anthropic a “supply-chain risk to national security” after it refused to allow mass domestic surveillance or fully autonomous weapons. On August 27, 2026, a federal court largely granted summary judgment for Anthropic, finding the actions retaliatory. (Risk factors · PDF p.83)
• Switched off. In June 2026 the U.S. Department of Commerce imposed worldwide export restrictions on Fable 5 and Mythos 5. Anthropic disabled both models for all customers until the restrictions were lifted. (Risk factors)
• $1.5B. The Bartz settlement: a separate line on the 2025 income statement, about a third of that year’s revenue. (Statements of operations)
• 22 banks. underwriters in the syndicate, led by Goldman Sachs and Morgan Stanley. The ticker will be ANT on the Nasdaq Global Select Market. (Underwriting · Cover)
• “A country of geniuses in a data center”. The prospectus’s own description of where this is going: millions of “Virtual Collaborators” running in parallel. (Prospectus summary · PDF p.20)
• 1,368%. Year-over-year revenue growth for the quarter ended June 30, 2026. Q2 revenue was $11.6B, against $16.3B for the whole first half. (Prospectus summary · MD&A)

Build it yourself

Everything above runs on bem’s public v3 API. Here is the shape of it:

# 1. One function per job (v3 API: extract / classify / split / join / enrich / parse / payload_shaping / render / send / evaluation)
curl -X POST https://api.bem.ai/v3/functions -H "x-api-key: $BEM_API_KEY" -d '{
  "functionName": "s1-financial-statements", "type": "extract", "enableBoundingBoxes": true,
  "outputSchemaName": "FinancialStatements", "outputSchema": { "type": "object", "properties": { "incomeStatement": { … } } } }'

# 2. Wire them into a graph — split item classes and classify choices become named edges
curl -X POST https://api.bem.ai/v3/workflows -H "x-api-key: $BEM_API_KEY" -d '{
  "name": "anthropic-s1-analyst", "mainNodeName": "classify_filing",
  "nodes": [ {"name": "classify_filing", "function": {"name": "s1-filing-classifier"}},
             {"name": "split_sections",  "function": {"name": "s1-section-splitter"}}, … ],
  "edges": [ {"sourceNodeName": "classify_filing", "destinationName": "s1_registration_statement", "destinationNodeName": "split_sections"},
             {"sourceNodeName": "split_sections", "destinationName": "financial_statements", "destinationNodeName": "financial_statements"},
             {"sourceNodeName": "financial_statements", "destinationNodeName": "judge_financials"}, … ] }'

# 3. Send the filing, then read every node's output from the trace
curl -X POST https://api.bem.ai/v3/workflows/anthropic-s1-analyst/call -H "x-api-key: $BEM_API_KEY" \
  -F file=@anthropic-s1.pdf -F callReferenceID=anthropic-s1-2026-10-26
curl https://api.bem.ai/v3/calls/$CALL_ID/trace -H "x-api-key: $BEM_API_KEY"

# 4. Browse the parsed corpus like a file system
curl -X POST https://api.bem.ai/v3/fs -H "x-api-key: $BEM_API_KEY" \
  -d '{"op": "grep", "pattern": "SpaceXAI", "context": {"bucket": "bkt_…"}}'

Point the same workflow at the next S-1, or at a 10-K, a credit agreement or a fund’s quarterly letters, and the desk runs again. When the S-1/A arrives with the price range filled in, the same calls will fill the blanks.

Source: Anthropic, PBC Form S-1, filed October 26, 2026 (SEC accession 0001628280-26-067168). Figures are as filed. Many offering figures are intentionally blank pending pricing. Nothing here is investment advice.