Claude Sonnet 5.5 Guide

Claude Sonnet 5.5 is Anthropic’s newest Sonnet model, released on September 28, 2026. It is the second Claude 5.5 model to launch, after Claude Opus 5.5.

It aims to balance intelligence, speed, and cost for the professional work you do every day.

Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per completed task.

Its strengths include coding, knowledge work, document creation, spreadsheets, presentations, visual understanding, computer use, and iterative workflows.

• Claude.ai
• Claude mobile apps
• Claude Code
• Claude API
• Amazon Web Services
• Google Cloud
• Microsoft Azure

API model ID

claude-sonnet-5-5

Prices per 1 million tokens:

Input tokens

$2

Output tokens

$10

Cache reads

$0.20

Cache writes

$2.50

Prompt caching can sharply cut the cost of repeated context. Batch processing can save up to 50% on eligible asynchronous workloads.

Use Sonnet 5.5 to:

• Navigate large codebases
• Ship new features
• Fix bugs
• Refactor code
• Write tests
• Review code
• Debug errors
• Create applications
• Build websites
• Edit across multiple files
• Run command-line tasks
• Iterate quickly on software

Use it to:

• Analyze reports
• Summarize research
• Compare documents
• Extract insights
• Synthesize large volumes of information
• Draft business analyses
• Review financial information
• Create executive summaries
• Analyze charts
• Turn raw information into finished deliverables

Use it for:

• Reports
• Proposals
• Strategy documents
• SOPs
• Research briefs
• Meeting documents
• Marketing plans
• Business plans
• Documentation
• Executive summaries

Sonnet 5.5 aims to write more clearly than earlier Sonnet generations, so drafts need less cleanup.

Use it to:

• Outline slide structures
• Follow existing templates
• Draft slide copy
• Turn reports into decks
• Build executive presentations
• Strengthen presentation narratives
• Polish slide content
• Analyze source material before you build slides

Use it to:

• Analyze spreadsheet data
• Structure new spreadsheets
• Create formulas
• Clean data
• Find patterns
• Explain calculations
• Run business analysis
• Build financial models
• Produce spreadsheet-based reports

Sonnet 5.5 can reason about visual inputs, including:

• Screenshots
• Charts
• Graphs
• Interfaces
• Documents
• Images
• Dashboards
• Visual layouts

Anthropic reports large gains in chart recognition and long-horizon visual tasks.

Sonnet 5.5 can support agentic workflows that operate computer interfaces. Possible workflows include:

• Browser tasks
• Research
• Data entry
• Application workflows
• Business operations
• Testing
• Information collection
• Multi-step digital processes

Keep permissions, verification, and human oversight in place before any computer-use system takes a consequential action.

Sonnet 5.5 brings improved design capabilities. Use it for:

• Landing pages
• Web interfaces
• Dashboards
• Frontend components
• UI improvements
• Visual hierarchy
• Layouts
• Responsive interfaces
• Design iterations
• Presentation design

Share screenshots, references, design requirements, and constraints to get stronger results.

Sonnet 5.5 can carry multi-step tasks that involve:

• Planning
• Reasoning
• Tool usage
• Code execution
• File manipulation
• Research
• Iteration
• Verification
• Computer interaction

Before you ship a production agent, define its permissions, stopping conditions, validation rules, and escalation points.

Higher effort can lift performance on hard tasks, but it can also raise latency and cost. More effort is not always better.

Anthropic reports these Sonnet 5.5 results:

Terminal-Bench 4.0

70.6% at Max effort

FrontierCode 1.1 Main

52.1% at Xhigh effort, 46.2% at Max effort

CursorBench 4.0

55.5%

GDPval-AA v2.1

1844

AA-Briefcase v1.1

1811

Humanity’s Last Exam with tools

64.5%

OSWorld 2.1

80.1% partial

Chartography

61.6% without tools

Results depend on methodology, effort settings, tools, scaffolding, and other conditions. Do not read them as universal real-world performance scores.

A strong prompt can include these parts.

Role: You are an expert [ROLE]. Objective: Your goal is to [OBJECTIVE]. Context: [CONTEXT] Inputs: [INPUTS] Requirements: [REQUIREMENTS] Constraints: [CONSTRAINTS] Process: [IMPORTANT STEPS OR CHECKS] Output: Produce [EXACT DELIVERABLE]. Quality criteria: Make the result accurate, practical, specific, logically consistent, and ready to use. Before finalizing: Check your work against every requirement, then fix any important omissions or inconsistencies.

• Be specific
• Describe the outcome you want, not vague instructions
• Supply relevant context
• Name the intended audience
• Set the output format
• Share examples when style or structure matters
• Keep instructions clearly apart from source material
• Spell out important constraints
• Split very complex projects into logical stages when it helps
• Give Claude the information it needs to solve the task
• Ask it to separate facts, assumptions, and uncertainties
• Request verification when accuracy matters
• Structure long prompts into sections
• Remove contradictory instructions
• Raise reasoning effort only when the task benefits
• Review important outputs before you act on them

• Build and debug software
• Analyze business documents
• Synthesize research
• Create presentations
• Analyze spreadsheets
• Plan marketing campaigns
• Build websites and interfaces
• Automate repetitive workflows
• Analyze screenshots, charts, and other visuals
• Run AI agents and internal business tools

• Audience research
• Customer research
• Competitor analysis
• Content strategy
• SEO planning
• Ad concepts
• Landing-page copy
• Email campaigns
• Social content
• Positioning
• Campaign analysis
• Customer feedback analysis
• Content repurposing
• Marketing reports

• Market research
• Business models
• Product strategy
• Customer discovery analysis
• Competitive research
• Pitch preparation
• Financial analysis
• SOP creation
• Hiring materials
• Product requirements
• Growth experiments
• Decision-support documents

• Feature development
• Debugging
• Codebase exploration
• Code review
• Testing
• Refactoring
• Documentation
• Architecture analysis
• Frontend development
• Terminal tasks
• Repository-wide changes
• Agentic coding workflows

• Content research
• Scripts
• Newsletters
• Social posts
• Content repurposing
• Idea generation
• Research synthesis
• Content calendars
• Audience analysis
• Digital products
• Course material
• Visual-content planning

• Reports
• Documents
• Presentations
• Spreadsheets
• Research
• Support workflows
• Operations
• Data analysis
• Internal knowledge
• Process documentation
• Workflow automation

Choose Sonnet 5.5 when you need:

• Strong capability
• Fast responses
• Lower API costs
• High-volume professional work
• Coding
• Document creation
• Design work
• Business analysis
• Repeated AI workflows
• Agentic tasks with contained complexity

Consider Opus 5.5 for tasks that involve:

• Very high complexity
• Open-ended problems
• Sustained judgment
• Difficult reasoning
• Complex agentic workflows
• Cases where maximum capability matters more than minimum cost

• Use lower effort for simple work
• Cache repeated context with prompt caching
• Batch work where it fits
• Stop resending irrelevant context
• Keep reusable instructions stable
• Route simple tasks to lower-cost workflows
• Save higher effort for difficult tasks
• Measure cost per completed task, not only cost per token
• Track input tokens, output tokens, tool calls, retries, latency, and success rate

• Share the actual source material
• Explain the business objective
• Define what success looks like
• Show examples of strong outputs
• Set constraints early
• Ask for evidence behind factual claims
• Require Claude to state uncertainty rather than guess
• Iterate on creative work
• Set explicit acceptance criteria for coding
• Give agents clear tool permissions
• Require verification before consequential actions
• Keep humans in high-impact decisions

Claude Sonnet 5.5 may still:

• Make factual mistakes
• Misinterpret ambiguous instructions
• Generate incorrect code
• Make calculation errors
• Misread incomplete context
• Produce plausible claims without support
• Fail partway through long multi-step workflows
• Use tools incorrectly
• Overlook edge cases

Verify AI-generated output before you use it for financial, legal, medical, security, business-critical, or other consequential decisions.

• Sonnet 5.5 adds stronger safeguards for certain high-risk cybersecurity requests
• Anthropic says routine software development and ordinary bug fixing are still supported
• Biology safeguards target certain high-risk requests
• Protections aim to reduce large-scale model distillation and reasoning extraction

• Define the outcome
• Give Claude the context it needs
• Attach relevant files, data, images, or code
• State the constraints
• Define the deliverable
• Pick a fitting effort level
• Let Claude complete the task
• Review factual claims and key decisions
• Give targeted feedback
• Iterate until the output meets your acceptance criteria

Need quick everyday professional work?

Use Sonnet 5.5.

Working on code, docs, decks, or spreadsheets?

Use Sonnet 5.5.

Running high-volume workflows on a budget?

Use Sonnet 5.5 at lower or medium effort.

Doing routine summaries or classification?

Use lower effort.

Doing content creation or standard coding?

Use medium effort.

Facing complex debugging or architecture?

Use higher effort.

Need repeated context for less?

Use prompt caching.

Running async work for up to 50% savings?

Use batch processing.

Tackling very complex, open-ended work?

Consider Opus 5.5.