Claude Sonnet 4.5
Anthropic's coding and agentic Claude model, released in 2025.
Claude Sonnet 4.5 is a general-purpose language model from Anthropic, released in 2025 and aimed at coding, reasoning and long-running agent tasks. The company says it is its most capable and most aligned frontier model so far, with strong results on software and computer-use evaluations.
- Scores 61.4% on OSWorld, up from 42.2% for Sonnet 4 four months earlier
- Anthropic reports more than 30 hours of focus on complex multi-step tasks
- Priced the same as Sonnet 4 at $3/$15 per million tokens
- Released under AI Safety Level 3 protections with CBRN classifiers
- Shipped with the Claude Agent SDK, the infrastructure behind Claude Code
What it is
Claude Sonnet 4.5 is a language model built by Anthropic and released in 2025. It sits in the Claude product line and is aimed at writing and reasoning about code, running software tools, and carrying out long multi-step tasks on a computer. Anthropic describes it as its strongest model for building agents and its best at computer use, with gains in reasoning and mathematics as well. It is offered through the Claude API under the name claude-sonnet-4-5, and it also powers the Claude apps and Claude Code.
How it came about
The model arrived alongside upgrades to Anthropic's own products. Claude Code gained checkpoints that save progress and let a user roll back, a refreshed terminal interface, and a native VS Code extension. The Claude API picked up context editing and a memory tool so agents can run longer. Anthropic also released the Claude Agent SDK, the infrastructure behind Claude Code, so developers can build their own agents with it. The company says it spent more than six months shipping Claude Code updates before opening that infrastructure up.
Inside the model
Anthropic has not published layer counts, parameter totals or training data for this model. It describes the design work as centred on agents: memory across long-running tasks, permission systems that balance autonomy with user control, and coordination between subagents working toward a shared goal. The model supports a context editing feature and a memory tool through the API. It is released under the company's AI Safety Level 3 protections, which add classifiers aimed at spotting dangerous inputs and outputs, especially ones related to chemical, biological, radiological and nuclear weapons.
What it can do
Anthropic reports the model leading the SWE-bench Verified software coding evaluation and holding focus for more than 30 hours on complex multi-step tasks. On OSWorld, a test of real-world computer work, it scores 61.4%, up from 42.2% for Sonnet 4 four months earlier. Experts in finance, law, medicine and STEM found better domain knowledge and reasoning than older models including Opus 4.1. Early customers reported figures such as a 44% cut in vulnerability intake time and an 18% rise in planning performance, though these are company-reported results.
How to use it
The model is available through the Claude API as claude-sonnet-4-5, and in the Claude apps and Claude Code. Pricing is unchanged from Claude Sonnet 4 at $3 per million input tokens and $15 per million output tokens. The Claude for Chrome extension became available to Max users from the waitlist. The Claude Agent SDK gives developers the same infrastructure that runs Claude Code.
Safety and testing
Anthropic calls this its most aligned frontier model, reporting less sycophancy, deception, power-seeking and encouragement of delusional thinking, plus better defence against prompt injection for agentic and computer use. The system card includes safety and alignment evaluations using techniques from mechanistic interpretability for the first time. ASL-3 classifiers can wrongly flag ordinary content; the company says false positives have fallen tenfold since first described and halved since Opus 4 in May, and interrupted chats can continue on Sonnet 4.
Reception and debate
Early customers quoted by Anthropic praised coding performance, long-context work and agent behaviour. Cursor said developers turn to it for their hardest problems, GitHub Copilot reported gains in multi-step reasoning and code comprehension, and Canva and Figma described improvements in design and prototyping work. One firm reported moving from a 9% error rate on Sonnet 4 to 0% on its internal code editing benchmark. These are customer statements published by the company rather than independent reviews.
Why it matters
Anthropic released the Claude Agent SDK alongside the model, opening up the infrastructure behind Claude Code so others can build agents for tasks beyond coding. The company frames the release as a step for agents that manage memory, permissions and subagents over long runs. Its computer-use score on OSWorld rose sharply over four months, and the model is positioned as the default upgrade path from Sonnet 4 at the same price.
Questions people ask about Claude Sonnet 4.5
What is Claude Sonnet 4.5 for?
Anthropic positions it for writing and reasoning about code, building agents, and driving a computer through real tasks. The company reports strong results on software benchmarks and computer-use tests, and says it handles long multi-step work with sustained focus.
How do I get it?
Developers call it as claude-sonnet-4-5 through the Claude API. It also runs inside the Claude apps and Claude Code. Pricing matches Claude Sonnet 4 at $3 per million input tokens and $15 per million output tokens.
How large is the model?
Anthropic has not published a parameter count, training data size or compute figure for Claude Sonnet 4.5. The company describes its capability and alignment work but not the underlying scale.
What safety measures come with it?
It ships under Anthropic's AI Safety Level 3 framework, with classifiers screening inputs and outputs for chemical, biological, radiological and nuclear weapons content. The system card covers alignment evaluations, including mechanistic interpretability techniques.
What changed from Sonnet 4?
Anthropic reports gains in coding, reasoning, mathematics, domain knowledge and computer use, plus lower rates of behaviours such as deception and sycophancy. The OSWorld computer-use score rose from 42.2% to 61.4% over four months.
Good reading and sources
There is no English Wikipedia article on Claude Sonnet 4.5 yet, so the facts on this page come from the sources listed above: anthropic.com/news/claude-sonnet-4-5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.