Models

A Claude AI model is one of Anthropic’s Opus, Sonnet, or Haiku systems: start with Sonnet 4.6 for most work, move simple high-volume tasks to Haiku 4.5, and reserve Opus 4.7 for the hardest reasoning, coding, and long-context analysis; use our Claude AI guide for the broader ecosystem.

Table of contents

• The current Claude lineup
• How the families differ
• Context windows and output limits
• What each model costs
• Older versions and version history
• Which model should you use?
• Honest take
• FAQ
• Sources

• Independent guide · c-ai.chat is not Anthropic.
• Official product · use claude.ai for Claude itself.
• Three main families · Opus, Sonnet, and Haiku.
• API billing · priced per million input and output tokens.

The current Claude lineup

The lineup has three main choices. Opus is the flagship tier. Sonnet is the default for most users and applications. Haiku is the fast, low-cost tier for simple tasks. Anthropic lists official model details in its model overview and publishes pricing at claude.com/pricing.

Model choice affects quality, latency, and cost. It also changes how you design prompts, how much context you send, and whether you route tasks between models. If you are choosing Claude by product capability rather than model capability, use our Claude features guide with this page.

Family	Current model	Primary role	API price	Context and output note
Opus	Claude Opus 4.7	Flagship quality for difficult work	$5/M input · $25/M output tokens	1,000,000-token context
Sonnet	Claude Sonnet 4.6	Recommended default	$3/M input · $15/M output tokens	1,000,000-token context · 128,000-token max output
Haiku	Claude Haiku 4.5	Fast, low-cost automation	$1/M input · $5/M output tokens	Best when the task is short, structured, and repeated

How the families differ

Opus, Sonnet, and Haiku are separate capability and cost tiers. Opus gives the most headroom for hard reasoning, nuanced writing, complex code work, and multi-document synthesis. Test Opus when the inputs are messy, the answer must be careful, or the cost of a wrong answer is high.

Sonnet is the practical center of the lineup. It is strong enough for most business writing, research support, software development, data analysis, and agent workflows, while costing less than Opus in the API. Most teams should start with Sonnet, measure failures, and move only the hard cases upward.

Haiku is the speed and cost model. Use it for classification, extraction, short summaries, routing, tagging, support triage, and other repeated tasks with clear instructions. Do not use Haiku as the first choice for subtle analysis, difficult architecture decisions, or prompts with many competing constraints.

The common mistake is treating the flagship model as the default for every task. That can hide weak prompt design and raise costs without improving outcomes. A better pattern is Sonnet for the main workflow, Haiku for simple sub-tasks, and Opus for escalation.

Context windows and output limits

A context window is the amount of material Claude can consider in one request: instructions, chat history, file text, retrieved passages, tool output, and response budget. In the current lineup covered here, Opus 4.7 and Sonnet 4.6 have 1,000,000-token context. Sonnet 4.6 also lists a 128,000-token max output.

Long context helps with large codebases, contracts, research sets, logs, and document collections. It is not a substitute for structure. If you paste an entire archive into a prompt, Claude may still miss details, over-weight irrelevant sections, or spend tokens on material that should have been filtered first.

• Use long context when the model must compare many sections at once.
• Use retrieval when only a small part of a large corpus is relevant.
• Ask for citations or section references when accuracy depends on source text.
• Keep output instructions tight, because output tokens are the expensive side of most requests.
• Test shorter prompts against long-context prompts before assuming bigger is better.

What each model costs

Claude API pricing is based on input tokens and output tokens. Input tokens are the text, files, retrieved context, tool results, and instructions you send to the model. Output tokens are the text the model generates. Anthropic publishes official pricing at claude.com/pricing and in the Claude platform pricing docs.

For subscriptions, annual billing, team plans, and API billing basics, use our Claude pricing guide. If you are building with model IDs, request bodies, prompt caching, or streaming, use our Claude API docs guide with Anthropic’s official docs.

Model	Input price	Output price	Cost profile	Practical use
Claude Opus 4.7	$5/M tokens	$25/M tokens	Highest cost in the current lineup	Reserve for difficult work, final review, high-stakes reasoning, and complex code tasks.
Claude Sonnet 4.6	$3/M tokens	$15/M tokens	Middle cost, strong quality	Use as the default for most professional and developer workflows.
Claude Haiku 4.5	$1/M tokens	$5/M tokens	Lowest cost in the current lineup	Use for fast, repeated, simple, or background tasks.

Claude web plans and API rates are separate. A claude.ai subscription controls product access and usage allowances. API usage is billed by model and token volume unless your account terms say otherwise.

Plan	Published price	Model-choice note
Free	$0	Good for trying Claude, not for estimating API spend.
Pro	$20/mo or $17/mo with annual billing	Personal plan pricing is separate from API token pricing.
Max	From $100/mo	Higher product access does not remove API token costs.
Team Standard	$25/seat or $20/seat with annual billing	Team plan pricing is separate from model API rates.
Team Premium	$125/seat or $100/seat with annual billing	Use official pricing for plan features and limits.
Enterprise	$20/seat base + API rates	Budget for both the seat base and model usage.

Prompt caching is most useful when many requests share the same long instructions, reference documents, schemas, or tool descriptions. It applies to cached input tokens, not every token. Batch API pricing can reduce costs for eligible non-urgent workloads, but it is not a fit for real-time chat.

Optimization lever	What it reduces	When to use it	Limitation
Prompt caching	Cached input tokens	Repeated prompts with shared instructions, policies, reference text, or schemas	It applies to cached input, not all tokens.
Batch API	Input and output tokens on eligible batch jobs	Offline jobs, bulk summarization, data processing, and non-urgent automation	It is not for real-time chat or interactive flows.
Model routing	Unnecessary use of larger models	Workflows with easy cases, hard cases, and clear escalation rules	It requires evaluation so routing does not lower quality.
Shorter outputs	Output tokens	Reports, summaries, extraction, and structured answers	Too much compression can remove useful detail.

Older versions and version history

Claude model names combine a family name with a version. Older docs, saved workflows, vendor integrations, and example code may still mention previous model IDs. Check Anthropic’s official model overview before copying an old ID into production.

Reference	How to treat it	Action for users
Opus 4.7	Current flagship model	Evaluate for the hardest tasks, long-context work, and final review.
Opus 4.6	Previous flagship at $5/M input and $25/M output tokens	Keep only when a pinned workflow still performs well; compare with Opus 4.7 before new work.
Sonnet 4.6	Current default model for most work	Start here for most API and claude.ai workflows.
Haiku 4.5	Current speed and cost model	Use for clear, repeated, low-risk automation.
Claude 4.0 or 3.x references	Older examples or legacy integrations	Check deprecation notes, migration guidance, and test results before relying on them.

Deprecation policy matters for production systems. Anthropic can retire older model IDs, change aliases, or recommend migration paths. Avoid silent model changes. Pin deliberate versions where appropriate, monitor official notices, and test migrations before switching high-value workflows.

Which model should you use?

The safest rule is simple: start with Claude Sonnet 4.6, move simple repeated work down to Claude Haiku 4.5, and move difficult or high-stakes work up to Claude Opus 4.7.

1. Start with Sonnet.

   Use Sonnet 4.6 for the first version of a workflow. Measure answer quality, latency, and cost before changing models.

2. Move simple sub-tasks to Haiku.

   Route classification, tagging, extraction, and short summaries to Haiku 4.5 when the instructions are clear and the risk is low.

3. Escalate hard cases to Opus.

   Send prompts to Opus 4.7 when Sonnet struggles with reasoning, ambiguity, multi-file code review, or high-stakes synthesis.

4. Use real prompts for testing.

   Benchmarks help, but your own data matters more. Test the prompts, files, tools, and edge cases your users will actually send.

5. Review the model mix regularly.

   Usage patterns change. Recheck routing rules, cached prompts, output length, and model quality after major workflow changes.

Workload	Start with	Move up when	Move down when
General business writing	Sonnet 4.6	The output needs more nuance, strategy, or final-review quality.	The task is template-based or only needs short transformations.
Software development	Sonnet 4.6	The task involves multi-file reasoning, difficult bugs, or architecture trade-offs.	The task is simple formatting, extraction, or code classification.
Claude Code workflows	Sonnet 4.6	The session needs more reasoning across a large codebase.	The task is simple explanation or repeated boilerplate support.
Document analysis	Sonnet 4.6	The corpus is large, conflicting, or high risk.	The task is extraction into a fixed schema.
Support triage	Haiku 4.5	Customer intent is ambiguous or escalation quality matters.	The task is simple tagging or routing.
Research synthesis	Sonnet 4.6	The source set is long, dense, or decision-critical.	The request only needs short summaries of known material.

For coding-heavy work, model choice is only part of the decision. Tool access, repository context, terminal behavior, and review flow also matter. Use our Claude resources for practical workflows and implementation notes.

Honest take

Do not start with the largest Claude AI model by default. Ask which model is reliable enough for the task. For most professional use, that answer is Sonnet 4.6. Use Haiku 4.5 for simple volume work. Use Opus 4.7 when extra capability changes the outcome.

The lineup is clear, but model names do not replace testing. Real prompts, real files, and real failure cases will tell you more than a label. For quick questions about accounts, limits, and plan choices, use our Claude FAQ.

FAQ

What is a Claude AI model?

A Claude AI model is one of Anthropic’s Claude model options. The main families are Opus, Sonnet, and Haiku. Each has a different balance of capability, speed, and API cost.

Which Claude model is best for most users?

Claude Sonnet 4.6 is the best default for most users. It balances reasoning quality, writing ability, coding support, speed, and cost better than starting every task on Opus or Haiku.

Is Claude Opus 4.7 better than Sonnet 4.6?

Opus 4.7 is the flagship model and is the stronger choice for the hardest prompts. Sonnet 4.6 is the better default for most work because it costs less in the API and is strong enough for many professional tasks.

What is Claude Haiku 4.5 for?

Haiku 4.5 is for fast, low-cost tasks. Common examples include classification, extraction, routing, tagging, support triage, and short summaries. Use Sonnet or Opus when the task needs more reasoning or nuance.

How much do Claude models cost in the API?

Claude Opus 4.7 costs $5/M input tokens and $25/M output tokens. Claude Sonnet 4.6 costs $3/M input tokens and $15/M output tokens. Claude Haiku 4.5 costs $1/M input tokens and $5/M output tokens. Check the official pricing page before budgeting.

Which Claude models support 1,000,000-token context?

This guide lists 1,000,000-token context for Claude Opus 4.7 and Claude Sonnet 4.6. Sonnet 4.6 also lists a 128,000-token max output. Use Anthropic’s official model overview for current model IDs, limits, and availability notes.

Are older Claude model IDs safe to use in production?

They can be, but only if you track Anthropic’s deprecation policy and test migrations. Do not rely on an old model ID just because an example or integration still shows it.

Are Claude plans and API pricing the same?

No. Claude plans control product access and usage allowances. API pricing bills by input and output tokens. Enterprise pricing includes a $20/seat base plus API rates.

Sources

• Anthropic company site
• Claude official product
• Claude pricing
• Claude model overview
• Claude platform pricing docs
• Prompt caching docs
• Batch processing docs

Independent guide. Not affiliated with Anthropic. For the official Claude product, visit claude.ai.