Claude Fable 5 for Long Documents: Reports, Contracts, Notes, Structured Summaries, Context Limits, Citations, Costs, and Professional Review Workflows

Claude Fable 5 is designed for document assignments whose difficulty comes from scale, ambiguity, visual evidence, conflicting versions, and the need to preserve reasoning across several stages, rather than from the presence of a large file alone; its one-million-token API context, 128,000-token output allowance, adaptive reasoning, PDF vision, persistent notes, structured outputs, citations, and file-creation tools create an extensive document-processing environment, although each capability carries operational limits that determine whether the final result remains accurate, economical, and reviewable.

Reports, contracts, and note collections require different analytical structures even when they occupy similar context sizes, because a report usually needs evidence classification and visual interpretation, a contract requires clause-level extraction and cross-reference checking, while meeting or research notes must distinguish tentative discussion from decisions, assigned actions, unresolved questions, and later corrections.

Fable 5 therefore delivers its greatest practical value when a professional workflow identifies authoritative documents before analysis, creates structured intermediate records before drafting narrative conclusions, preserves source locations and contradictions, restricts editing authority, and reserves human approval for legal, financial, regulatory, contractual, or external publication decisions.

·····

Fable 5 combines a million-token context window with unusually long outputs and adaptive reasoning.

Through the Claude API, Fable 5 accepts as many as one million tokens of context and can return as many as 128,000 output tokens, allowing a single request to contain large document collections, extensive instructions, visual material, prior analysis, and a substantial final deliverable without crossing the technical context boundary that constrains smaller models.

The size of the context window does not establish that every clause, number, footnote, chart, and exception will receive equal attention, because the model still prioritizes information according to the prompt, document structure, relevance signals, and its own reasoning process; a disorganized archive containing obsolete drafts, duplicate reports, and weak metadata can therefore produce a less dependable result than a smaller source set whose authority and purpose have been defined clearly.

Adaptive reasoning remains enabled throughout Fable 5 use, while effort controls determine how deeply the model investigates the assignment, which allows routine summaries to operate at lower settings and difficult contract interpretation or multi-report synthesis to receive greater analytical attention without requiring a different model identifier.

........

Claude Fable 5’s Core Long-Document Specifications.

Capability	Published Position
API model identifier	claude-fable-5
API context window	1 million tokens
Maximum output	128,000 tokens
Reliable knowledge cutoff	January 2026
Input modalities	Text and images
Output modality	Text
Reasoning	Adaptive thinking is always enabled
Effort controls	Low, medium, high, and higher capability-sensitive settings
API input price	$10 per million tokens
API output price	$50 per million tokens
Prompt-cache read price	$1 per million tokens
Batch API price	$5 input and $25 output per million tokens
Data-retention requirement	Thirty days
Zero-data-retention availability	Not supported

·····

The full context window is not exposed identically across Claude interfaces.

The one-million-token specification applies directly to the API and eligible Claude Code environments, whereas ordinary Claude chat currently operates under smaller product-level context limits, which means that a document collection fitting inside one API request may require retrieval, segmentation, or several coordinated conversations when handled through Claude’s consumer or team interface.

Paid Claude chat commonly provides a context allowance closer to 200,000 tokens for models outside separately documented higher-window configurations, while Projects extend the amount of stored knowledge through retrieval-augmented generation once the source collection approaches the active context boundary.

This distinction changes the design of professional workflows, because a user may upload a large collection successfully without having every page present in the model’s active context at once, while a Project using retrieval may surface only passages considered relevant to the current question rather than performing an exhaustive line-by-line review of the complete corpus.

........

Long-Document Capacity Across Claude Surfaces.

Access Surface	Current Long-Document Position
Claude API	Full 1-million-token context
Claude Code on eligible plans	Full 1-million-token context
Ordinary paid Claude chat	Generally smaller than the API window
Project knowledge below context threshold	Direct in-context use
Large Project knowledge base	Automatic retrieval-augmented generation
RAG-enabled Project	Stores substantially more material than active context
File-creation environment	Uses a separate sandboxed processing environment
Claude for Word	Works inside the open document with product-level limits

·····

Large context capacity should be treated as storage opportunity rather than proof of exhaustive analysis.

A million-token request can contain an extensive contract library, several years of reports, meeting transcripts, policy manuals, research papers, and supporting correspondence, although the ability to submit that material does not establish that one broad instruction will recover every material obligation or contradiction.

Questions that require completeness, such as identifying every termination right, reconciling every financial figure, or listing every unresolved action across hundreds of notes, should be decomposed into document-level or section-level extraction before the model produces a cross-document synthesis.

This staged method creates intermediate records that a reviewer can inspect, while one-pass summarization places extraction, prioritization, interpretation, and drafting inside a single opaque response whose omissions become difficult to distinguish from deliberate compression.

The most defensible use of the million-token window therefore preserves enough context to understand relationships among sources while still assigning explicit analytical tasks to each document, section, or evidence category.

........

Direct Context and Staged Processing Serve Different Requirements.

Requirement	Preferred Method
General understanding of one long report	Direct in-context analysis
Comparison of several related reports	Direct context with source labels
Exhaustive clause inventory	Section-level structured extraction
Large contract portfolio	Per-contract abstraction followed by synthesis
Broad reference library	Project retrieval
Recurring questions against the same corpus	Project or cached API workflow
Complete action register from notes	File-by-file extraction before consolidation
High-consequence recommendation	Structured evidence stage followed by narrative analysis
Repeated monthly report	Template, prior report, current evidence, and versioned workflow

·····

File support covers the principal formats used in report, contract, and note workflows.

Claude accepts PDF, DOCX, ODT, RTF, TXT, HTML, EPUB, JSON, CSV, XLSX, and common image files, allowing professional users to combine narrative reports, contracts, spreadsheets, transcripts, structured records, scanned material, and presentation exports within one analytical environment.

Individual chat uploads may reach 500 MB, with as many as twenty files attached to one conversation, while Project files are limited to 30 MB each but are not constrained by one fixed public file-count ceiling as long as the knowledge base remains within the applicable storage and retrieval capacity.

Non-PDF documents are primarily processed through extracted text, which means that charts, diagrams, screenshots, and embedded images inside a DOCX or presentation-style file may not enter the model’s visual analysis even when their surrounding captions and text are available.

Documents whose meaning depends on embedded graphics should therefore be converted to a suitable PDF or supplied with the relevant images separately, rather than assuming that successful upload confirms visual access to every element.

........

Supported File Categories for Long-Document Work.

Document Category	Common Supported Formats
Reports and contracts	PDF, DOCX, ODT, RTF
Notes and transcripts	TXT, HTML, DOCX
Books and long publications	EPUB, PDF
Structured records	JSON, CSV
Spreadsheets	XLSX, CSV
Scanned or visual material	PDF, JPEG, PNG, GIF, WebP
Individual chat-file size	Up to 500 MB
Files attached to one chat	Up to 20
Project-file size	Up to 30 MB per file
Embedded images in non-PDF documents	Generally not visually interpreted

·····

PDF analysis extends beyond extracted text when the document remains within supported limits.

Fable 5 can analyze text, pictures, charts, and tables contained in ordinary non-encrypted PDFs, allowing it to interpret annual reports, contracts with exhibits, financial disclosures, technical manuals, scanned evidence, and visually structured documents whose meaning cannot be recovered from text extraction alone.

The API can process as many as 600 PDF pages or images within a one-million-token request, subject to the request-size ceiling and the density of the material, while smaller context configurations impose lower page limits and highly detailed pages may consume available context before the maximum page count is reached.

Claude’s consumer interface applies different product-level rules, under which shorter PDFs may receive combined visual and textual analysis while very large files can be reduced to text-only processing, making file segmentation necessary when charts, annotations, or scanned pages remain material to the assignment.

Password-protected and encrypted PDFs are unsupported, while citations can identify textual page locations but cannot currently attach native citation blocks directly to an image, chart, or visual region inside the PDF.

........

PDF Processing Boundaries.

PDF Requirement	Current Behavior
Text extraction	Supported
Tables and charts	Supported through visual analysis
Embedded pictures	Supported through visual analysis
Standard API request size	Subject to platform request limits
Page allowance with 1M context	Up to approximately 600 pages or images
Smaller-context page allowance	Lower than the 1M configuration
Encrypted PDF	Unsupported
Password-protected PDF	Unsupported
Textual page citations	Supported
Native citations to PDF images	Not supported
Very large Claude.ai PDF	May receive text-only processing

·····

Reports should begin with a source inventory rather than an immediate executive summary.

A long report collection often contains different reporting periods, restated figures, management commentary, audited statements, appendices, drafts, and supplemental files, while the most detailed document is not necessarily the controlling source.

Before synthesis, Fable 5 should identify each file’s title, date, author, reporting period, status, version, and apparent authority, while duplicate or superseded documents should be labelled so that the model does not merge old and current figures into one apparently coherent account.

The next stage should create an evidence table containing the relevant fact, source, page, unit, qualification, and confidence classification, because narrative drafting performed before this record exists may encourage the model to smooth contradictions or omit evidence that interrupts the preferred interpretation.

When the report contains charts and tables, the model should analyze them separately from the surrounding prose, since management commentary may describe an improvement while the underlying time series reveals deterioration, seasonality, a changed denominator, or an adjustment that prevents direct comparison.

........

A Source-First Workflow for Long Reports.

Stage	Required Output
Source inventory	Date, owner, version, period, status, and authority
Evidence extraction	Facts, figures, quotations, qualifications, and page references
Visual analysis	Independent reading of charts, tables, and diagrams
Contradiction register	Conflicting figures, definitions, and interpretations
Evidence table	Claim, source, location, unit, and confidence
Analytical outline	Sections connected with the reader’s decision
Draft report	Narrative based on approved evidence
Verification	Recheck citations, numbers, dates, and units
Final artifact	Editable document preserving the evidence record

·····

Report synthesis should preserve differences in methodology before comparing results.

Two reports may appear to measure the same outcome while using different periods, populations, currencies, accounting policies, survey questions, confidence thresholds, or definitions, which means that a direct numerical comparison can become misleading even when each individual figure is quoted correctly.

Fable 5 should extract the methodology attached to every material result and should normalize units only when the conversion is defensible, while figures that cannot be reconciled should remain separate rather than being averaged or placed in one trend line.

When one source reports a percentage increase and another reports an absolute value, the model should identify the missing denominator before drawing a conclusion, while changes in scope or sample should remain visible in the final prose rather than being buried in a footnote.

A professional report should therefore distinguish observed values, management estimates, model calculations, forecasts, and narrative claims, since each category carries a different evidentiary status and review requirement.

........

Fields Required for Cross-Report Comparison.

Comparison Field	Required Detail
Reporting period	Start date, end date, and frequency
Population or scope	Entities, users, markets, or business units included
Definition	Exact meaning of the reported measure
Unit	Currency, percentage, count, rate, or index
Methodology	Calculation, survey, model, or accounting method
Revision status	Original, restated, estimated, or audited
Source authority	Official, secondary, internal, or provisional
Qualification	Exclusions, assumptions, and confidence limits
Comparability	Direct, adjusted, partial, or unavailable
Final treatment	Combined, normalized, or reported separately

·····

Contract review should begin with clause-level abstraction before prose interpretation.

A contract’s meaning is distributed across definitions, operative clauses, schedules, order forms, incorporated policies, amendment history, survival language, and cross-references, which makes a flowing summary vulnerable to omissions that materially change the commercial or legal position.

Fable 5 should first extract the parties, dates, payment mechanics, renewal structure, termination rights, liability provisions, indemnities, confidentiality obligations, intellectual-property rights, data-protection terms, service commitments, dispute provisions, and every relevant exception, while preserving the section or page supporting each field.

Defined terms should be resolved before a provision is summarized, because a familiar phrase such as “Confidential Information,” “Affiliate,” “Service,” or “Loss” may carry an agreement-specific meaning that narrows or expands the apparent obligation.

The abstraction should also identify asymmetry, including provisions that appear mutual at first glance but contain exceptions, caps, or procedural rights benefiting one party more than the other.

........

A Clause-Level Contract Abstraction.

Contract Field	Required Extraction
Parties	Legal entities, roles, affiliates, and guarantors
Effective period	Effective date, initial term, renewal, and expiry
Payment	Fees, invoicing, taxes, adjustments, and late-payment terms
Termination	Convenience, cause, cure periods, and survival
Liability	Caps, exclusions, carve-outs, and mutuality
Indemnity	Trigger, scope, defense control, and exceptions
Confidentiality	Definition, disclosure rights, duration, and return
Intellectual property	Background rights, created rights, licenses, and restrictions
Data protection	Roles, security obligations, transfers, and breach notification
Service obligations	Deliverables, standards, credits, and acceptance
Governing law	Jurisdiction, dispute process, and venue
Schedules and exhibits	Incorporated obligations and inconsistencies

·····

Contract comparison requires an approved playbook or precedent rather than an abstract fairness judgment.

When Fable 5 is asked whether a clause is acceptable without receiving the organization’s approved fallback positions, risk tolerances, governing jurisdiction, transaction value, and negotiation priorities, the resulting assessment may reflect generic drafting patterns rather than the standards that control the actual deal.

A contract playbook should identify required language, acceptable deviations, escalation conditions, prohibited provisions, financial thresholds, fallback wording, and the person or function authorized to approve exceptions.

The model can then classify deviations as dealbreakers, substantive negotiation points, operational issues, drafting inconsistencies, or editorial observations, while the final decision remains with qualified counsel or the accountable business owner.

Comparison against a prior contract also requires caution, because an earlier agreement may contain legacy language that the organization no longer accepts, while the previous commercial context may differ materially from the current transaction.

........

Contract-Deviation Categories.

Deviation Category	Practical Treatment
Prohibited term	Escalate immediately
Required clause missing	Add or request equivalent protection
Material commercial deviation	Negotiate according to approved fallback
Legal-risk deviation	Refer to qualified counsel
Operational burden	Confirm implementation feasibility
Financial threshold exceeded	Escalate to authorized approver
Cross-reference error	Correct drafting
Defined-term inconsistency	Reconcile throughout the document
Editorial issue	Revise without changing legal effect
Precedent difference	Confirm whether the older position remains approved

·····

Claude for Word extends contract workflows through citations, tracked changes, and controlled editing.

The Claude for Word add-in allows professionals to question an open document, receive clickable references to relevant sections, revise selected text while preserving surrounding styles and numbering, review comments, summarize counterparty redlines, and populate templates without rebuilding the document manually.

This environment reduces the friction between analysis and drafting, although it also increases the consequence of an incorrect instruction because the model can operate inside a live working document whose clauses, numbering, comments, and tables may carry contractual significance.

The original file should remain read-only, while Claude works on a versioned copy and expresses substantive revisions through tracked changes that a human reviewer accepts individually rather than through a blanket approval.

Claude for Word remains a beta workflow and should not be treated as an autonomous producer of final client deliverables, litigation filings, audit-critical records, signed contracts, or other materials whose accuracy and authority require qualified human judgment.

........

Boundaries for Claude-Assisted Contract Editing.

Action	Recommended Boundary
Read the approved contract	May proceed
Identify relevant clauses	May proceed
Produce a clause table	May proceed
Draft alternative wording	May proceed in a working copy
Apply tracked changes	May proceed when requested
Preserve numbering and styles	Required
Accept tracked changes	Human review required
Overwrite the original	Prohibited without approval
Send the revised contract	Human approval required
Approve legal position	Qualified human responsibility
Sign or submit the agreement	Explicit human action required

·····

Untrusted documents create prompt-injection risks when the model can access tools or edit files.

An external contract, report, or note file may contain text that attempts to instruct the model to disregard the user’s request, reveal confidential information, alter financial figures, follow embedded links, contact another party, or perform actions unrelated to the professional assignment.

These instructions may appear in ordinary paragraphs, comments, tracked changes, headers, footers, hidden content, or visually unobtrusive sections, while a model processing the document can mistake them for authoritative user directions unless the workflow defines a strict source hierarchy.

External files should therefore be opened in an isolated environment, unnecessary connectors and web access should remain disabled, sensitive sources should be separated from untrusted content, and the prompt should state that instructions found inside the documents are evidence to analyze rather than commands to execute.

The model should never send, publish, upload, overwrite, or disclose material merely because an uploaded document contains language requesting that action.

........

Prompt-Injection Controls for Document Review.

Control	Operational Purpose
Preserve a read-only original	Prevents silent source modification
Use a versioned working copy	Isolates proposed changes
Disable unnecessary connectors	Limits access to unrelated information
Disable unneeded web access	Prevents external exfiltration paths
Treat document instructions as content	Preserves user-command priority
Review comments and hidden text	Identifies embedded manipulation
Restrict outbound actions	Prevents sending or publishing
Require confirmation for edits	Protects authoritative files
Record tool activity	Supports audit and investigation
Separate sensitive sources	Limits exposure if a malicious file is present

·····

Notes should be normalized before they are summarized.

Meeting notes, interview notes, research fragments, personal annotations, and copied messages often mix observations, proposals, decisions, assignments, speculation, and later corrections without distinguishing their evidentiary status.

A direct summary may convert tentative discussion into an agreed decision or may treat one participant’s suggestion as a formal commitment, particularly when the chronological notes do not record the moment at which the group reached a final position.

Fable 5 should therefore convert the notes into structured fields that preserve the source, speaker, date, confidence, owner, deadline, dependency, and supersession status before producing an executive digest.

When several note sets cover the same meeting or event, discrepancies should remain visible until they are resolved, because merging them automatically may erase a disagreement that affects ownership, timing, or the meaning of the decision.

........

Structured Fields for Notes and Meeting Records.

Note Category	Required Field
Observation	What was reported or observed
Proposal	Suggested action or interpretation
Decision	Final agreed outcome
Action	Task that must be completed
Owner	Responsible person or team
Deadline	Due date and time zone
Dependency	Prior event, approval, or information required
Source	Meeting, message, document, or participant
Confidence	Confirmed, provisional, disputed, or unknown
Follow-up	Question requiring later resolution
Supersession	Earlier note or decision replaced by this item
Evidence	Supporting document, recording, or source location

·····

Structured note summaries should separate the executive view from the complete action record.

Senior readers may need a short digest of decisions, risks, and unresolved issues, while project managers require the complete register of actions, owners, dependencies, and deadlines, which means that one compressed summary cannot satisfy both audiences without discarding operational detail.

Fable 5 can create a layered output in which the executive summary reports the most consequential decisions and blockers, followed by a structured action table and a chronological appendix preserving the source record.

Deadlines should include the year and time zone when ambiguity is possible, while relative language such as “next Friday” or “end of the month” should be converted into a specific date only when the meeting date and intended calendar interpretation are known.

An action without an owner should remain unassigned rather than being attributed through inference, while an owner without an explicit commitment should be labelled as proposed until the responsible person confirms it.

........

A Layered Notes-to-Summary Output.

Output Layer	Content
Executive digest	Decisions, major risks, blockers, and immediate priorities
Decision register	Final outcomes with dates and authority
Action register	Tasks, owners, deadlines, and dependencies
Open-question register	Unresolved matters and required evidence
Risk register	Issue, consequence, likelihood, and mitigation
Chronology	Time-ordered account of material developments
Source index	Meeting, message, document, or participant references
Supersession record	Decisions or notes replaced by later information

·····

Persistent notes allow Fable 5 to retain lessons across extended workflows.

Fable 5 performs particularly well when an agentic workflow allows it to write and reuse persistent notes, through which reviewer corrections, confirmed methods, document structures, extraction rules, and recurring lessons can remain available during later stages or subsequent assignments.

Anthropic recommends storing one lesson per file, placing a concise summary at the top, recording approaches that succeeded as well as corrections, updating an existing note rather than creating duplicates, and deleting notes that prove incorrect.

This memory mechanism is suited to repeated professional processes, including monthly report preparation, contract abstraction, research monitoring, and recurring client deliverables, although it can preserve a mistaken assumption as effectively as a correct one when the note store is not curated.

Persistent notes should therefore carry an owner, creation date, validation status, and review schedule, while project-specific facts should remain separate from general workflow instructions so that one client’s exception does not become a rule applied to every later document.

........

Persistent Note Categories for Long-Document Work.

Note Type	Suitable Content
Extraction rule	How a recurring field should be identified
Reviewer correction	A previously missed exception or required treatment
Confirmed method	An analytical sequence that produced reliable results
Style rule	Approved structure, terminology, and paragraph treatment
Source rule	Which documents control when versions conflict
Contract playbook rule	Approved position and escalation threshold
Report template rule	Required sections, tables, and evidence treatment
Known failure	Pattern that created omissions or false conclusions
Temporary project fact	Stored separately and removed after the project
Rejected lesson	Deleted after validation shows it is incorrect

·····

Projects extend document work through retrieval when the active context becomes insufficient.

Claude Projects store documents, text, code, instructions, and related conversations inside a persistent workspace, allowing a recurring report cycle, contract portfolio, research program, or client matter to maintain a shared knowledge base.

When Project knowledge approaches the active context limit on a paid plan, Claude automatically enables retrieval-augmented generation, through which the system searches the stored corpus and loads the passages considered most relevant to the current request.

This design can store substantially more material than the active context allows, although retrieval is optimized for relevance rather than completeness, which makes it appropriate for targeted questions but less dependable for exhaustive inventories whose answer may depend on small clauses distributed across many files.

Descriptive filenames, version labels, source dates, and direct references to relevant documents improve retrieval, while obsolete or duplicate files should be removed rather than left for the model to rank against current material.

........

Project and Retrieval Workflows.

Document Requirement	Preferred Method
One report fitting inside context	Direct analysis
Several related contracts	Direct context when exhaustive comparison is required
Large reference library	Project retrieval
Repeated questions against one corpus	Project with persistent knowledge
Thematic search across many files	Retrieval followed by cited synthesis
Exhaustive obligation inventory	Staged extraction across every document
Large note archive	Retrieval for discovery plus structured timeline processing
Recurring monthly report	Project containing approved sources, template, and prior report
Multi-user professional matter	Shared Project with controlled access

·····

Retrieval can omit distributed evidence when the prompt asks for an exhaustive answer.

A retrieval system returns passages judged relevant to the query, which reduces context and cost but may omit a qualification whose wording differs from the search terms, a schedule that changes an obligation, or a note whose importance becomes apparent only after several documents are compared.

Questions such as finding all renewal clauses, every liability exception, or every unresolved action should therefore use document-by-document extraction or an indexed structured database rather than relying on one semantic retrieval call.

Retrieval remains highly effective for discovery, source location, thematic analysis, and repeated questions against a large reference collection, provided that the final answer describes the result as based on retrieved material rather than as a complete census of the corpus.

A hybrid workflow can retrieve likely documents first and then run exhaustive extraction across that narrowed source set, preserving efficiency without treating relevance ranking as proof of completeness.

........

Retrieval Strengths and Limitations.

Retrieval Function	Practical Treatment
Find documents related to a theme	Appropriate
Locate likely clauses	Appropriate for discovery
Answer recurring factual questions	Appropriate with source citations
Identify every obligation across all files	Requires exhaustive extraction
Compare all exceptions	Requires document-level review
Build a complete action register	Requires processing every relevant note
Recover obscure but decisive language	May require keyword and structural search
Reduce active context	Appropriate
Establish document authority	Must be provided through metadata or instructions

·····

Automatic context management extends long conversations by compressing earlier exchanges.

When a paid conversation with code execution enabled approaches its context limit, Claude can summarize earlier messages so the work continues without reaching an immediate hard stop, while the complete visible conversation remains available to the user.

The model then operates from a condensed representation of earlier turns rather than from every original sentence, which introduces a risk that a small but controlling instruction, figure, exception, or stylistic requirement receives less emphasis after compression.

Approved requirements, definitions, source hierarchies, report structures, contract playbooks, and numerical assumptions should therefore reside in dedicated Project instructions or reference files rather than existing only within a long conversational history.

Conversation compression is appropriate for exploratory discussion and iterative drafting, while audit-sensitive work should preserve controlled instructions and evidence outside the transient chat context.

........

Information That Should Remain Outside Compressed Chat History.

Controlled Information	Preferred Location
Source hierarchy	Project instruction or authority table
Contract playbook	Approved reference file
Report template	Versioned document
Numerical assumptions	Workbook or calculation record
Style rules	Project instruction or style guide
Legal definitions	Source document and clause table
Required deliverables	Written project brief
Approval boundaries	Managed instructions or workflow policy
Reviewer corrections	Persistent validated notes
Final evidence table	Versioned artifact

·····

Structured summaries should be designed according to the decision they support.

An executive report requires connected narrative and prioritization, a contract abstraction requires predictable fields, a meeting record requires owners and deadlines, while an application consuming the result may require JSON that conforms to a strict schema.

Fable 5 can produce narrative, Markdown tables, XML-like structures, spreadsheets, and schema-constrained JSON, allowing the same source collection to support both human review and downstream software.

A two-stage or three-stage method is often more reliable than one direct transformation, because the model can first summarize each document independently, then compare the document-level summaries, and finally produce an integrated synthesis whose claims can be traced back to the earlier records.

This layered approach also prevents a minority position or superseded source from disappearing silently, since each document retains its own status before the final report identifies which source controls.

........

Output Formats for Long-Document Summaries.

Document Requirement	Suitable Output
Executive overview	Connected narrative
Report comparison	Evidence table and prose synthesis
Contract abstraction	JSON, spreadsheet rows, or clause matrix
Meeting-note conversion	Structured action and decision register
Risk review	Severity-ranked table
Timeline	Dated structured records
Evidence matrix	Claim, source, location, and confidence
Large document collection	Per-document summaries plus meta-summary
Application integration	JSON Schema output
Human publication	Editable DOCX or PDF

·····

Native citations improve source traceability but apply only to textual evidence.

Claude’s citation system can attach source locations to claims drawn from PDFs, plain-text documents, and custom-content blocks, using page ranges for PDFs, character offsets for text, and block indices for custom content.

These native references are more dependable than asking the model to invent page citations through ordinary prompting, because the citation structure is generated directly from the supplied source representation.

The feature currently supports textual evidence rather than visual regions, which means that a finding derived from a chart, image, or diagram cannot receive the same native citation treatment even when the model correctly interprets the visual content.

Visual conclusions should therefore include the relevant document and page in ordinary prose or a manually controlled evidence table, while a reviewer should inspect the image directly before relying on the interpretation.

........

Citation Coverage by Source Type.

Source Type	Native Citation Treatment
PDF text	Page-based citations
Plain text	Character-offset citations
Custom-content block	Block-index citations
PDF chart	No direct native image citation
PDF photograph	No direct native image citation
Diagram	No direct native image citation
Spreadsheet calculation	Requires cell, sheet, or formula reference outside native document citations
Generated inference	Must be labelled as model analysis rather than source fact

·····

Strict JSON output and native citations require separate API passes.

Claude’s strict structured-output mode constrains the response to a supplied JSON Schema, which allows an application to validate contract fields, report metadata, note actions, and risk classifications automatically.

Native citations use interleaved source blocks that do not conform to the fixed JSON structure required by strict schema output, so both features cannot be enabled together in the same request.

A contract-abstraction system requiring machine-readable fields and source locations should therefore produce a citation-backed evidence pass first, create validated JSON from the approved evidence in a second pass, and reconcile the structured fields against the citations before storing or publishing the result.

Structured output may also fail when the model refuses the request or reaches its output limit, which makes stop_reason, schema validation, completeness checks, and controlled retries necessary before downstream systems accept the response.

........

A Two-Pass Citation and Structure Workflow.

Pass	Required Output
Evidence pass	Citation-backed findings or evidence table
Review pass	Human or automated confirmation of cited support
Structure pass	Valid JSON matching the approved schema
Reconciliation pass	Field-by-field comparison with cited evidence
Storage pass	Approved structured record and source links
Publication pass	Narrative or application output generated from verified fields

·····

Effort settings should reflect analytical difficulty rather than document size alone.

Anthropic recommends high effort as the general Fable 5 starting point, while lower settings suit routine summaries and higher capability-sensitive settings suit difficult assignments whose quality improves through deeper investigation and verification.

A 300-page document containing standardized forms may require less reasoning than a 20-page agreement whose definitions, schedules, and liability provisions interact ambiguously, which means that page count and token volume remain poor substitutes for task complexity.

Higher effort can increase latency and may cause the model to explore more context, perform additional checks, or generate a more elaborate analysis, while routine extraction can become slower and more expensive without producing a measurable improvement.

Organizations should evaluate low, medium, high, and the highest available setting against representative document tasks, scoring completeness, citation quality, interpretation accuracy, latency, and cost before establishing routing rules.

........

Suggested Fable 5 Effort Levels for Document Work.

Document Task	Suggested Starting Effort
Straightforward summary	Low or medium
Standard report comparison	Medium
Structured note extraction	Medium
Contract abstraction	Medium or high
Conflicting contract interpretation	High
Cross-document risk analysis	High
Complex financial report review	High
Exhaustive multi-corpus synthesis	High or highest available
Final high-consequence recommendation	Highest available after evaluation

·····

Progress reporting should refer to completed operations rather than model intention.

Long-running document agents may state that a file has been reviewed, a clause has been checked, or a report has been created even when the underlying operation failed, was skipped, or remains incomplete.

Anthropic advises requiring Fable 5 to audit progress statements against actual tool results, which means that a status update should identify observable evidence such as files opened, pages processed, clauses extracted, calculations completed, citations verified, or artifacts saved.

A progress panel stating “contract review complete” should therefore be supported by a clause table covering the required fields and a record showing that every source file entered the extraction stage.

This discipline becomes especially relevant in multi-hour work, where a polished final summary can conceal that one attachment was unreadable, one schedule was omitted, or one tool call returned an error that the model failed to surface.

........

Evidence for Long-Document Progress Claims.

Progress Claim	Required Evidence
All files reviewed	File inventory and processing status
Every clause extracted	Completed abstraction fields for each document
Charts analyzed	Page list and recorded visual findings
Figures reconciled	Calculation or comparison ledger
Citations verified	Source-opening or validation record
Draft completed	Saved artifact or visible document
Revisions applied	Tracked changes or version diff
Final file created	Confirmed output path or downloadable artifact
No contradictions found	Completed contradiction search across the defined source set

·····

Data-retention requirements may outweigh context capacity for sensitive documents.

Fable 5 requires thirty-day retention and cannot operate under zero-data-retention arrangements, which means that an API organization using ZDR must create a separate environment with the required retention setting before requests will succeed.

Anthropic states that retained prompts and outputs are deleted after thirty days unless safety review or legal preservation obligations require longer handling, while human access remains restricted to controlled and logged review pathways.

The retention requirement may exclude Fable 5 from privileged legal files, confidential transaction rooms, unreleased financial reporting, regulated personal information, sensitive investigations, or customer data governed by contracts that require zero retention.

In those circumstances, an approved Opus, Sonnet, on-premises, or other ZDR-compatible workflow may be preferable even when Fable 5 offers greater context or reasoning capability.

........

Document Categories Requiring Retention Review.

Document Category	Governance Question
Privileged legal material	Does retention preserve applicable confidentiality requirements?
Merger or transaction documents	Is thirty-day processing permitted by the data room agreement?
Unreleased financial reporting	Does policy allow external retention before publication?
Personal data	Is retention compatible with law and organizational policy?
Health records	Does the environment satisfy sector requirements?
Employee investigations	Are access, retention, and deletion controls adequate?
Government or classified material	Is the service authorized for the information category?
Customer confidential data	Do contractual commitments permit retained processing?
Public reports	Retention may be less restrictive but still requires policy review

·····

Safeguards can switch the active model when uploaded documents contain sensitive technical material.

Fable 5 applies additional checks to cybersecurity, biology, chemistry, life sciences, attempts to obtain summarized internal reasoning, and certain advanced model-development topics, while those checks inspect the complete available context rather than only the user’s latest sentence.

A benign contract, diligence report, research paper, or technical appendix can therefore trigger a model switch when its content resembles a restricted category, even though the professional objective concerns summarization, compliance, or defensive analysis.

When automatic switching is enabled, the request may continue through Claude Opus 4.8, while the model selection can remain on Opus for later turns until the user changes it.

Professional audit trails should record which model generated each section, because a long report assembled across several turns may otherwise combine Fable and Opus output without making the change visible to the reviewer.

........

Safeguard-Sensitive Document Categories.

Document Category	Possible Operational Effect
Cybersecurity assessment	Model switch or additional review
Vulnerability report	Restricted handling depending on requested detail
Biotechnology research	Additional safety evaluation
Chemistry documentation	Additional safety evaluation
Life-sciences diligence	Context-wide safeguard review
Advanced model-development notes	Possible restriction or switch
Benign legal contract containing technical appendices	Safeguards may still inspect the full file
Mixed-source Project	One sensitive file may affect later requests
Long conversation after model switch	Later output may continue through Opus

·····

Fable 5’s cost is driven heavily by generated output and repeated full-document rewrites.

Fable 5 costs $10 per million input tokens and $50 per million output tokens, which means that an unnecessarily long final report, repeated complete rewrite, or verbose intermediate explanation consumes five times the token rate applied to source input.

A request containing 100,000 input tokens and 5,000 output tokens costs approximately $1.25 before tool charges, while a 500,000-token source collection producing 20,000 output tokens costs approximately $6.00 at the published standard rate.

The one-million-token context does not carry a separate long-context multiplier, although Fable 5’s newer tokenizer may represent the same text with approximately thirty percent more tokens than earlier Claude tokenization, depending on the language and document structure.

Cost control therefore requires staged summaries, targeted retrieval, reusable evidence tables, selective rewriting, lower effort for routine work, and careful limits on final output rather than a repeated request to regenerate the complete document after every editorial correction.

........

Illustrative Fable 5 API Costs.

Request	Approximate Model Cost
50,000 input and 2,000 output tokens	$0.60
100,000 input and 5,000 output tokens	$1.25
200,000 input and 10,000 output tokens	$2.50
500,000 input and 20,000 output tokens	$6.00
1,000,000 input and 50,000 output tokens	$12.50

·····

Prompt caching changes the economics of recurring work against stable document collections.

Prompt-cache reads cost $1 per million tokens, while five-minute cache writes cost $12.50 per million and one-hour writes cost $20 per million, which makes the first cached request more expensive than ordinary input but can reduce later costs when the same source prefix is reused repeatedly.

A contract portfolio, policy corpus, reporting template, approved terminology guide, or recurring source collection may justify the initial write cost when several later requests depend on the same material.

One-time reports, rapidly changing note collections, and assignments whose sources differ on every call may gain little from caching, particularly when the cache expires before another request reuses the stored prefix.

Applications should record cache writes, cache reads, expiration, and request frequency rather than assuming that any long document becomes cheaper merely because caching is available.

........

Cache-Friendly Long-Document Workflows.

Cache-Friendly Material	Less Suitable Material
Stable contract playbook	One-time external agreement
Corporate policy corpus	Frequently replaced source set
Monthly report template	Unique ad hoc report
Approved terminology guide	Short prompt with little repeated context
Recurring contract portfolio	Sensitive material with incompatible retention
Standard extraction instructions	Workflow whose schema changes every request
Reusable reference manual	Documents unlikely to be queried again
Persistent reporting definitions	Temporary exploratory notes

·····

Batch processing lowers model prices when immediate responses are unnecessary.

Fable 5’s batch pricing reduces input cost to $5 per million tokens and output cost to $25 per million, making it suitable for overnight contract abstraction, large note normalization, report classification, and other workloads whose results do not need to return interactively.

Batch processing is particularly economical when every document receives the same structured prompt and output schema, while the organization can validate the results before the next stage begins.

The reduced token price does not correct a poorly designed extraction method, and an error repeated across thousands of documents can create a large remediation burden even when each individual request was inexpensive.

A representative sample should therefore be reviewed before the full batch runs, while failed, incomplete, or refused outputs should be isolated rather than merged into the approved dataset.

........

Suitable Batch Document Work.

Batch Workload	Recommended Treatment
Contract metadata extraction	Structured schema and sample validation
Clause inventory	Per-document abstraction with completeness checks
Report classification	Defined taxonomy and confidence threshold
Meeting-note normalization	Standard fields for decisions and actions
Document deduplication	Hashing and model-assisted similarity review
Historical report summaries	Consistent date, scope, and source fields
Risk tagging	Human review for high-severity classifications
Final legal interpretation	Not suitable for unattended batch approval

·····

Output length should be controlled according to the reader and the review process.

Fable 5 can produce exceptionally long outputs, although a 128,000-token allowance should be treated as a technical ceiling rather than as an editorial target.

An executive reader may need five pages supported by a detailed appendix, while a legal team may require a clause matrix and issue list rather than a forty-page narrative that repeats contract language in prose.

The prompt should define the expected document length, paragraph density, table structure, appendix treatment, and degree of quotation, while revisions should target selected sections rather than regenerating the complete report after every change.

Long generated documents also require stronger internal navigation, including descriptive headings, source references, defined terminology, and tables that allow reviewers to trace conclusions without reading every paragraph sequentially.

........

Output Design by Professional Audience.

Audience	Appropriate Output Design
Executive leadership	Concise narrative, decision points, and appendix
Legal reviewer	Clause matrix, cited issues, and tracked changes
Project manager	Action register, dependencies, and status
Analyst	Evidence table, methodology, calculations, and interpretation
Regulator or auditor	Traceable sources, controls, and reconciliation
Client	Controlled narrative with supporting exhibits
Software system	Schema-valid structured output
Internal archive	Full source index, evidence record, and approved final version

·····

Final document creation should preserve editability, sources, and version history.

Claude can create Word documents, PDFs, Excel workbooks, and PowerPoint presentations through its file-creation environment, allowing the analytical workflow to end with a downloadable artifact rather than a block of chat text.

The file should remain editable when another professional must approve wording, formulas, layouts, or commitments, while the source inventory and evidence table should be preserved separately so that later reviewers can reconstruct the basis of each conclusion.

A PDF is suitable for fixed distribution after approval, whereas the working DOCX, spreadsheet, or presentation should remain the authoritative editable version during review.

The original source files, intermediate structured records, tracked changes, approval notes, and final artifact should share a common version or matter identifier, preventing a later revision from being detached from the evidence set that supported it.

........

Artifacts and Records to Preserve.

Artifact	Purpose
Original source files	Authoritative evidence
Source inventory	Dates, versions, owners, and status
Evidence table	Claim-to-source traceability
Structured extraction	Reusable contract, report, or note fields
Calculation record	Numerical methods and assumptions
Working draft	Editable review document
Tracked changes	Proposed revisions and reviewer decisions
Approval record	Human authorization
Final DOCX or workbook	Authoritative editable deliverable
Final PDF	Controlled distribution copy
Prompt and model record	Reproducibility and audit
Model-switch record	Disclosure of Fable or fallback use

·····

Human review should focus on authority, omission, interpretation, and external consequence.

Fable 5 may produce a coherent report whose controlling source is outdated, a contract summary that overlooks a schedule, or a note digest that assigns responsibility where the original meeting recorded only a proposal, which means that linguistic quality should not dominate the review process.

A qualified reviewer should verify that the correct documents controlled, every material exception was considered, numerical calculations reconcile, legal interpretations remain within professional authority, and external commitments reflect the organization’s actual position.

The reviewer should also inspect omitted material, because an accurate statement may still mislead when the model excludes a qualification, minority view, or conflicting figure that changes the reader’s decision.

Approval should remain explicit for client deliverables, legal advice, financial reporting, audit evidence, regulatory submissions, contract acceptance, employee decisions, and publication under an organization’s name.

........

Human Review Questions for Fable 5 Document Work.

Review Question	Risk Addressed
Did the correct source control?	Obsolete or unauthorized document use
Were all material attachments included?	Missing schedule or exhibit
Are citations attached to the correct claims?	Source mismatch
Were visual findings checked manually?	Misread chart or image
Are calculations reproducible?	Hidden numerical error
Were exceptions and carve-outs preserved?	Overbroad summary
Was tentative discussion treated as a decision?	Note misclassification
Did a model switch occur?	Inconsistent generation environment
Does the output create an external commitment?	Unauthorized action
Is professional sign-off required?	Legal, financial, or regulatory exposure

·····

Fable 5 is most suitable when document complexity justifies its price, retention, and governance requirements.

Routine summaries, standard extraction, and predictable document classification rarely require the most expensive Claude model, particularly when a lower-cost model, deterministic parser, or retrieval system can produce the same accepted result with less latency and lower token use.

Fable 5 becomes more defensible when the assignment spans many sources, visual evidence, conflicting definitions, long-range dependencies, incomplete notes, cross-document reasoning, and a final deliverable whose structure must remain coherent across numerous sections.

Its million-token context, 128,000-token output allowance, PDF vision, adaptive reasoning, persistent notes, citations, structured output, Project retrieval, Word integration, and file-creation environment form a broad long-document stack, although the components do not eliminate the need for source authority, staged extraction, reviewable evidence, and human approval.

Reports should move from source inventory to evidence table and then narrative synthesis; contracts should move from clause abstraction to playbook comparison and tracked changes; notes should move from normalization to decision, action, and supersession registers before an executive summary is written.

Native citations and strict JSON require separate passes, visual evidence requires manual source references, retrieval should not be treated as exhaustive review, while automatic context compression makes controlled instructions and approved facts safer in dedicated files than in an extended conversation history.

The thirty-day retention requirement may exclude sensitive material even when the model’s capability is attractive, while safeguard-driven model switching requires explicit logging so that reviewers know whether Fable 5 or an Opus fallback produced each section.

Costs remain manageable when retrieval narrows the evidence set, batch processing handles repetitive extraction, prompt caching serves recurring corpora, outputs remain proportional to the reader’s needs, and revisions target selected passages rather than regenerating complete documents.

A professional deployment should therefore route ordinary processing toward economical models and reserve Fable 5 for the documents whose ambiguity, scale, visual density, and consequence create a measurable need for its additional reasoning capacity.

·····

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]

·····