The paper trail

Small print.
Wide open.

A clear path from a claim to the page it came from. Here’s the filing, the research, and how we read it.

The documentS–1Anthropic, PBC26 OCT 2026 · 421 PP.
01The original SEC filingThe authoritative source, in your browser.SEC · HTML 02Download the saved filingThe complete 421-page document, optimized for downloading.PDF · 46 MB 03Download the full researchFindings, customer examples, calculations, and source references.Markdown

Page by page

Go straight to the source.

Page references count physical pages in the saved PDF, starting at 1. The printed filing number is shown separately when available.

Enter a page from 1 to 421. No PDF download needed.

From the filing

PDF page 210

Printed filing page 199 · Image transcription

View on the SEC
Read the page text
Table of Contents

The Claude Model Family

The Claude model family is currently composed of four model classes: Mythos (including Fable), Opus, Sonnet, 
and  Haiku.  Our  model  classes  can  be  used  individually  for  certain  tasks  or  in  combination  with  one  another. 
They  offer  broad  flexibility  to  choose  the  ideal  performance,  cost,  and  speed  for  each  use  case.  All  models 
support text, code, and multimodal inputs. 

• Mythos  (including  Fable).  Our  most  capable  model  class,  delivering  frontier  performance  across  all 
domains  and  solving  problems  AI  has  not  reliably  handled  before,  especially  in  coding  and  long-
running  agentic  tasks.  Given  its  advanced  capabilities  in  sensitive  areas  like  software  vulnerability 
discovery  and  biology,  we  deploy  this  underlying  model  in  two  distinct  configurations.  Fable  is 
generally  available  and  runs  with  our  production  safeguards,  which  block  or  limit  performance  on 
certain sensitive topics, including by routing requests to a less capable Claude model. Mythos runs with 
safeguards  calibrated  for  vetted  organizations  and  individuals  doing  cybersecurity  and  life  sciences 
work, and is available through our trusted-access programs, including our Cyber Verification Program. 
We first made the class available in April 2026 as Claude Mythos Preview through Project Glasswing, 
made Fable 5 generally available and offered Mythos 5 to Project Glasswing participants in June 2026, 
and released Mythos 5.1 and Fable 5.1 in September 2026. 

•

•

Opus.  Our  powerful  workhorse  model  class  for  reasoning-intensive  tasks.  Our  most  recent  Opus 
release, Claude Opus 5.5, is designed to serve as the enterprise daily driver, powering agents that run 
projects  end  to  end  and  advancing  capabilities  across  agentic  coding,  computer  use,  and  knowledge 
work. Opus models consistently rank among leading models on key industry benchmarks.

Sonnet.  Our  versatile  model  class  for  everyday  tasks,  providing  a  balance  of  performance,  cost,  and 
speed for the widest set of general-purpose use cases, including high-volume sub-agents in multi-agent 
orchestration setups. 

• Haiku. Our lowest-cost and fastest model class. Haiku models are designed for high-volume workloads 

like customer service agents and large-scale document processing.

199

[Embedded graphic 1 — OCR text]
Value
delivered to
customers

Platform

More productive employees
* Claude

Chat Collaborate with Claude
Code Build software with Claude
Cowork Delegate tasks to Claude

Faster processes
Do-it-yourself Work with us
with our developer by building custom
building blocks agents together

Embedded intelligence

Create Claude-fueled experiences

Tag Works alongside teams
SPECIALIZED PRODUCTS EXAMPLES EXAMPLES
F A Accelerate drug Knowyour Streamline Legal Financial Recruiting
Sesuiley wisest © r | & a discovery customer contract review casework agent analysis agent agent
Platform features Context management Infrastructure Harnesses
Extended thinking Adaptive thinking —_ Effort controls Model Context Protocol Prompt caching Agent SDK
Tooluse Longcontext Structured outputs Model Hardware Standard Batch processing Claude Managed Agents
Code execution Websearch Files API Agent skills Fast mode
“%€ClaudeMythos Claude Opus “%ClaudeSonnet Claude Haiku

Frontier intelligence

Powerful intelligence

General-purpose intelligence

Cost-effective intelligence

This page includes text transcribed from an image. Check the original for tables, images, and exact formatting.

Find a section 16 starting points

Section links may begin before the exact paragraph cited in a finding. PDF numbers below use physical pages.

  1. Founders’ letterp. 3SEC
  2. Defined termsp. 12SEC
  3. Prospectus summaryp. 16SEC
  4. Risk factorsp. 38SEC
  5. Management’s discussion & analysisp. 144SEC
  6. Businessp. 190SEC
  7. Governancep. 254SEC
  8. Executive compensationp. 275SEC
  9. Related-person transactionsp. 293SEC
  10. Principal & selling stockholdersp. 299SEC
  11. Capital-stock rightsp. 303SEC
  12. Shares eligible for future salep. 318SEC
  13. Underwritingp. 327SEC
  14. Auditor change & expertsp. 345SEC
  15. Financial statements & notesp. 347SEC
  16. Part II & exhibit indexp. 413SEC

How this was made

Read in full.
Kept in context.

A snapshot of the filing dated , rather than a live record of later events.

3primary reviewers
421pages reviewed
43image instances checked

The review

Three primary reviewers read non-overlapping ranges of the complete main filing. All 43 embedded image instances were transcribed and visually reviewed, including the founders’ letter, charts, and customer case studies. Financial tables were also checked visually for dates, units, and column alignment.

A separate reviewer checked the draft’s calculations, accounting interpretations, and customer qualifications. An additional check covered the later merged findings. The result is 62 findings and 21 customer examples.

The boundary

Companion contracts and other exhibits named in the filing’s index are separate documents and were not independently reviewed. Customer outcomes, technical benchmarks, forecasts, and scientific priority claims remain attributed company or customer disclosures.

The numbers

Dollar figures are U.S. dollars. H1 means the six months ended June 30; Q2 means the three months ended June 30. The 2026 interim financial statements are unaudited. Our calculations show their inputs and are labeled separately from reported values.

Blank fields stay undisclosed. Figures with different periods, conditions, or potentially overlapping scopes are kept separate. An earlier tender price is not an IPO price.

What the evidence labels mean
Reported in filing
A company disclosure, with the period and accounting basis retained.
Company estimate / company claim
An estimate or assertion attributed to the company, with its uncertainty preserved.
Our calculation
A derived figure using cited inputs. It is not a separately reported company metric.
Customer example
A selected case study, with its measurement limits. It is not representative ROI.
Proposed / conditional
A status qualifier for an arrangement that is not yet effective or depends on stated conditions.

About updates. This edition is tied to the October 26, 2026 filing. Any later revision should identify its new source date and preserve the basis of earlier claims.

The footnotes are only the beginning.Explore the findings Read the story
The original, in context

Open the footnote.

A closer look at the filing

Anthropic · Form S-1PDF 1

Opening page 1…

Filed October 26, 2026. PDF is approximately 46 MB and opens only when requested. Section links may precede the exact cited paragraph.