Claude Sonnet 5

Supported Functionality

Item	Specification
Input	Text, images
Output	Text
Context	1,000,000 tokens
Max Output	128,000 tokens; up to 300,000 in Message Batches API beta
Vision	✓ Supported
Function Calling	✓ Supported

Description

Claude Sonnet 5 is an Anthropic language model released on June 30, 2026, with the official Claude API ID claude-sonnet-5. It is designed to balance capability and response speed in development and enterprise workflows. The model accepts text and images, produces text, and offers a one-million-token context window with a standard 128,000-token maximum output.

Compared with Sonnet 4.6, it improves coding, tool use, agentic execution, and knowledge work while enabling adaptive thinking by default. On the Claude API and Google Cloud, it also supports newer browser and computer-use tools. Anthropic lists its comparative latency as fast, making it suitable for sustained execution where interaction speed matters.

Key Capabilities

• Agentic Coding: Carries a task from requirements through code changes, tests, and fixes across multiple steps.
• Tool-Driven Agents: Uses browsers, terminals, and other tools to advance workflows and adapt when intermediate results fail.
• Computer Use: Acts on interface state to click, type, and verify steps in existing business software.
• Reasoning and Knowledge Work: Analyzes business information, research material, and constrained problems into actionable conclusions.
• Long-Context Processing: Connects project documents, code, images, and prior actions within a one-million-token input window.
• Vision: Interprets screenshots, charts, and document images for interface work and data checking.
• Adjustable Thinking Depth: Uses adaptive thinking by default and varies reasoning through effort.

Technical Strengths

Feature	Benefit
Default Adaptive Thinking	Applies reasoning to complex tasks without manually assigning a fixed thinking budget to every request.
One-Million-Token Context	Holds large repositories, long documents, and work history for cross-source checks.
Improved Agentic Follow-Through	Completes more downstream steps in coding, tool use, and knowledge tasks than Sonnet 4.6.
Browser and Computer Tools	Interacts with websites and desktop interfaces on supported platforms, reducing custom integration work.
Adjustable effort	Tunes reasoning depth across short answers, routine execution, and difficult analysis.
Real-Time Safeguards	Provides identifiable refusals for high-risk cybersecurity requests so applications can handle them appropriately.

Frequently Asked Questions

What API settings should I change when moving from Sonnet 4.6 to Sonnet 5?

Update the official model ID and remove manual-thinking budget_tokens as well as non-default temperature, top_p, or top_k values; those settings produce a 400 error. Adaptive thinking is on by default, so retest whether max_tokens leaves enough room for both thinking and final text.

Does Sonnet 5's one-million-token window hold the same amount of text as Sonnet 4.6?

Not necessarily. Anthropic says the new tokenizer generally produces more tokens for identical text, so the same nominal window can hold less raw text; recount migrated prompts rather than reusing earlier estimates.

When should I raise Sonnet 5's effort level?

Test higher effort for multistep coding, long tool chains, or decisions requiring conflicting evidence to be checked. For classification, extraction, and short answers, first see whether a lower setting already meets quality and latency targets.

Is 300K output available for ordinary Sonnet 5 requests?

No. The standard maximum is 128K tokens; 300K is a beta limit for Anthropic's Message Batches API with its required beta configuration, not the default online output limit.

Should I use Sonnet 5 or Opus 5?

Start testing Sonnet 5 for everyday coding, automation, and faster interaction; compare Opus 5 when difficult reasoning or extended engineering and agent tasks fail acceptance criteria. Use identical tasks to assess accuracy, completion, recovery, and actual latency rather than choosing by tier name alone.