Claude Opus, Sonnet & Haiku Differences | Model Selection Decision Chart

Model Selection Decision Chart — Comparison Table of the Current 3 Models: Opus / Sonnet / Haiku

This comparison table shows the current three models as presented in the Anthropic official documentation "Models overview" (API prices are in USD excluding tax, as of 2026-04-23).

Source: Anthropic Models overview / Prices in USD excluding tax, as of 2026-04-23

Claude Opus 4.7 — Top-Tier Model ($5/$25 per MTok, 1M Tokens)

Overview — GA on 2026-04-16, Best Performance for Complex Reasoning and Agentic Coding

Claude Opus 4.7 became generally available on April 16, 2026. Anthropic positions it as "the most capable general-purpose model for complex reasoning and agentic coding." Source

Key New Features — Vision 2,576px / xhigh / Adaptive Thinking / Task Budget

High-Resolution Image Support — Up to 2,576px / 3.75MP with 1:1 Coordinate Mapping

Claude Opus 4.7 supports image input up to 2,576px / 3.75MP. This is more than three times the detail of the previous limit (1,568px / 1.15MP), enabling improved performance in computer use, screenshot analysis, and technical diagram interpretation. Coordinates now map 1:1 to actual pixels, eliminating the need to calculate scale factors. Source

xhigh Effort Level — New Starting Point for Coding Agents

The effort level controls how deeply the model "thinks" before responding. It ranges from standard (speed-first) to xhigh (deepest thinking); higher levels improve accuracy at the cost of more processing time and higher cost. A practical approach is to start with xhigh for difficult tasks and step down once you have sufficient results.

A new xhigh effort level has been added. The official documentation recommends using xhigh as the starting point for coding and agentic use cases, and at least high for most use cases where intelligence matters. Source

Adaptive Thinking — Replacement for Extended Thinking, Off by Default

Adaptive Thinking is a mechanism where the model automatically adjusts the depth of reasoning based on problem difficulty. It answers simple questions quickly while automatically spending more time on harder problems, reducing the need for developers to manually specify thinking budgets.

In Opus 4.7, Adaptive Thinking is enabled by setting thinking: {type: "adaptive"}. It is off by default. The previous Extended Thinking (budget_tokens specification) has been removed from Opus 4.7, and Adaptive Thinking is its replacement.

Task Budget (Public Beta) — Model Self-manages Token Spend Across Long Agent Loops

A feature now available in public beta allows you to specify a token spending guideline for an entire long-running agent loop. Unlike max_tokens (a hard cap per request), this is an advisory limit that the model itself tracks and self-regulates. Source

Benchmarks — Coding Performance Up 13–70% vs. Previous Generation

Key highlights from this section:

• Coding: 13–70% performance improvement over the previous generation across multiple benchmarks
• Agentic tasks: Significantly more consistent than Claude Opus 4.6 on complex multi-step workflows
• Vision capabilities: Major improvement in interpreting chemical structures and complex technical diagrams
• Knowledge worker tasks: Improvements in .docx change tracking, .pptx editing, and chart analysis

API Pricing — $5/$25 per MTok, Up to 90% Discount with Prompt Caching

Key highlights from this section:

Prompt caching offers up to 90% discount, and batch processing offers a 50% discount.

Claude Sonnet 4.6 — Balanced Standard Model ($3/$15 per MTok, 1M Tokens)

Overview — Released 2026-02-17, Best Balance of Speed and Intelligence

Claude Sonnet 4.6 is a hybrid reasoning model that provides "the best balance of speed and intelligence." Released February 17, 2026. It features a 1M token context window and combines high agentic intelligence with fast response times. Source

Key Features of Sonnet 4.6 — Both Extended and Adaptive Thinking, 1M Tokens

Key highlights from this section:

• Extended Thinking support: Choose between standard responses and extended thinking mode
• Adaptive Thinking support: Same adaptive reasoning as Opus 4.7
• 1M token context: Available in beta via API
• Multi-platform: Available on Claude.ai, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry

Primary Use Cases for Sonnet 4.6 — Multi-file Reasoning, Long-running Agents, Enterprise Workflows

Key highlights from this section:

• Complex reasoning across multi-file codebases
• Autonomous workflows for long-running agents
• Enterprise workflow automation via browser and computer use
• Finance, research, and content generation

Sonnet 4.6 API Pricing — $3/$15 per MTok, About 40% Cheaper Input Than Opus

Key highlights from this section:

Prompt caching offers up to 90% discount, and batch processing offers a 50% discount.

Choosing Between Opus 4.7 and Sonnet 4.6 — Opus for High Complexity, Sonnet for Cost Control

Opus 4.7 excels at agentic coding and high-resolution vision processing. When high intelligence is needed while keeping costs down, Sonnet 4.6 is the right choice. Sonnet 4.6's input price is about 40% cheaper, which provides a significant cost advantage for systems sending large volumes of requests.

Claude Haiku 4.5 — Lightweight Fast Model ($1/$5 per MTok, SWE-bench 73.3%)

Overview — Fastest and Cheapest, Maintains Near-Frontier Intelligence

Claude Haiku 4.5 is the fastest and most cost-efficient model. Anthropic positions it as having near-frontier intelligence, achieving 73.3% on SWE-bench Verified. Source

Key Features of Haiku 4.5 — Fastest Latency, Extended Thinking, Parallel Sub-Agent Support

Key highlights from this section:

• Fastest latency: Ideal for real-time applications
• Extended Thinking support: Access thinking capabilities while maintaining cost efficiency
• Multi-agent support: Parallel processing as a sub-agent
• 200k token context

Primary Use Cases for Haiku 4.5 — Real-time Response, High-volume Batch, Free-tier Agents

Key highlights from this section:

• Real-time chatbots and customer service
• High-volume batch processing and log classification
• Code generation and refactoring
• Financial analysis and market monitoring
• Parallel processing for research surveys
• AI agents for free-tier users

Haiku 4.5 API Pricing — $1/$5 per MTok, Cheapest Among Current Models

Key highlights from this section:

Prompt caching offers up to 90% discount, and batch processing offers a 50% discount.

Choosing Between Sonnet 4.6 and Haiku 4.5 — Haiku for Speed, Sonnet for Quality

For use cases where response speed is the top priority and costs need to be minimized, Haiku 4.5 is the right choice. For tasks requiring complex reasoning or long context, Sonnet 4.6 delivers better quality.

Model Selection Flow by Use Case — Quick Reference Table for 9 Use Cases

Key highlights from this section:

xhigh Effort Level Operating Guidelines — xhigh for Coding, Minimum high for Intelligence-critical Tasks

In Opus 4.7, you can adjust reasoning intensity via the effort field in output_config. The official documentation recommends the following guidelines:

• xhigh: Recommended starting point for coding and agentic use cases
• high: Minimum for most use cases where intelligence matters
• low / standard: For systems prioritizing speed and cost

The exact token consumption multiplier for each effort level is not detailed in the official documentation [unconfirmed]. Combining task_budget and max_tokens for cost management is the official recommendation. When using Managed Agents, effort settings are handled automatically.

Legacy Models — Sonnet 4 / Opus 4 Retire on 2026-06-15, Migration Required

Confirmed Retiring Models — Sonnet 4 / Opus 4 End on 2026-06-15

According to the Anthropic official documentation, the following two models are scheduled to retire on June 15, 2026:

• Claude Sonnet 4 (claude-sonnet-4-20250514): Migration to Claude Sonnet 4.6 recommended
• Claude Opus 4 (claude-opus-4-20250514): Migration to Claude Opus 4.7 recommended

Legacy Models Still Available — Opus 4.6 / 4.5 / 4.1, Sonnet 4.5

The following models are still available but migration to current models is recommended. Check official announcements for retirement timelines.

API Compatibility Between Models — 5 Breaking Changes to Watch When Migrating to Opus 4.7

Breaking changes that may affect you when migrating to Claude Opus 4.7 (Messages API only; no API impact for Managed Agents users).

For detailed migration steps, also refer to the Claude Opus 4.7 Migration Guide.

FAQ — 7 Common Questions About Model Selection

Here are answers to the questions we receive most frequently from readers.

Q: Which Claude model is the most intelligent?

As of 2026-04-23, Claude Opus 4.7 has the highest intelligence among current models. Anthropic positions it as "the most capable general-purpose model." Source

Q: Which Claude model is the cheapest?

Claude Haiku 4.5 is the cheapest. Input $1/MTok, output $5/MTok (USD excluding tax, as of 2026-04-23).

Q: For coding, which is better — Opus 4.7 or Sonnet 4.6?

For high-complexity, long-running autonomous agentic coding, Opus 4.7 is the better choice. For everyday coding tasks where cost is also a consideration, Sonnet 4.6 is more suitable.

Q: Can I still use Claude Opus 4.5?

Yes, it remains available as a legacy model. However, since the retirement schedule is updated periodically, please check the source for the latest information.

Q: Will my existing code work after migrating to Opus 4.7?

If your code specifies non-default values for temperature, top_p, or top_k, or uses budget_tokens, you will encounter 400 errors. We recommend reviewing the checklist in the Migration Guide before migrating.

Q: Which is newer — Claude 4.5 or Claude 4.6?

Claude Sonnet 4.6 is the newer model (Sonnet 4.5 is legacy). In the Opus line, models were released in the order Opus 4.5, 4.6, 4.7, with Opus 4.7 being the current top model.

Sources (Primary Information)

The following primary sources were directly referenced in creating this article. Always verify the latest accurate information at each link.

• Anthropic: Claude Models Overview — Current and legacy model tables, pricing, API IDs
• Anthropic: What's New in Claude Opus 4.7 — New features, breaking changes, Adaptive Thinking, xhigh effort level
• Anthropic: Claude Opus 4.7 Announcement Blog — Release overview, benchmarks
• Anthropic: Claude Opus Product Page
• Anthropic: Claude Sonnet Product Page
• Anthropic: Claude Haiku Product Page