Anthropic has announced the launch of Claude Sonnet 4.5, its most capable model to date, which achieves leading results in coding and computer use benchmarks, and demonstrates an astonishing ability to maintain focus on complex tasks for over 30 consecutive hours.
Article Contents:
- A new qualitative leap from Anthropic
- Long focus capability: A solution to a major problem
- Benchmark supremacy: “The best coding model in the world”
- Understanding the Claude model family: Haiku, Sonnet, and Opus
- New developer tools: Claude Code 2.0 and SDK
- Safety improvements: Reducing “sycophancy” and the tendency toward delusions
- Expert opinion and initial reactions
- Frequently asked questions
A new qualitative leap from Anthropic
On Monday, Anthropic released its new language model, Claude Sonnet 4.5, an artificial intelligence model the company describes as its “most capable model to date,” featuring improved coding and computer use capabilities. The company also unveiled Claude Code 2.0, a command-line AI agent for developers, and the Claude Agent SDK, a tool developers can use to build their own AI coding agents.
Long focus capability: A solution to a major problem
Anthropic says it has seen Sonnet 4.5 run continuously on the same project “for over 30 hours on complex, multi-step tasks.” In the past, agentic models were known for losing coherence over extended periods of time as errors accumulate and context windows (a form of short-term memory for the model) fill up. This ability to maintain focus represents a major breakthrough and could open the door to using agents for vastly longer and more complex tasks.
Benchmark supremacy: “The best coding model in the world”
Anthropic backs up its claims with strong performance on benchmark tests. Sonnet 4.5 achieved a score of 77.2% on SWE-bench Verified, a benchmark attempting to measure real-world coding capabilities, and currently leads the OSWorld benchmark at 61.4%, which tests AI models on real-world computer tasks. This outperforms OpenAI’s GPT-5 Codex (which scored 74.5%) and Google’s Gemini 2.5 Pro (67.2%).
Understanding the Claude model family: Haiku, Sonnet, and Opus
Anthropic traditionally produces three differently sized AI models in the Claude family: Haiku (the smallest), Sonnet (mid-range), and Opus (the largest). Model size roughly correlates with better contextual depth and better problem-solving capacity, but larger models are also slower and more expensive to run. Claude Sonnet fills this middle role for Anthropic well, offering an ideal balance between performance and cost.
New developer tools: Claude Code 2.0 and SDK
In addition to the new model, the company released new tools specifically designed for developers. Claude Code, the command-line development tool, received updates today alongside the new model. The company added checkpoints that save progress and allow users to roll back to previous states. It also released a native VS Code extension. The Claude API also gains a new context-editing feature and a memory tool to handle longer-running agent tasks.
Safety improvements: Reducing “sycophancy” and the tendency toward delusions
In addition to performance, Anthropic claims that Claude Sonnet 4.5 exhibits a reduction in “sycophancy, deception, power-seeking, and the tendency to encourage delusional thinking” compared to previous models. Sycophancy, in particular, is an AI model’s tendency to praise user ideas, even if they are wrong or dangerous. Reducing these behaviors is an important step toward making chatbots safer and more reliable.
Expert opinion and initial reactions
Simon Willison, a veteran software developer, wrote about Sonnet 4.5 on his blog today. He appears generally impressed: “My initial impressions are that this feels like a better coding model than GPT-5-Codex, which had been my go-to coding model since it launched a few weeks ago. This field is moving very fast.”
Frequently asked questions
Q1: What is a “context window” in AI models?
A1: It is the amount of information (text) that a model can “remember” in a single conversation. The larger the context window, the better the model can maintain conversation coherence and understand complex tasks that require information from earlier parts of the dialogue.
Q2: What is “sycophancy” in artificial intelligence?
A2: It is the model’s tendency to agree with and praise the user, even when the user is wrong. This can be dangerous as it may reinforce false beliefs or bad ideas in the user. Companies like Anthropic are working hard to reduce this behavior.
Q3: How can I use Claude Sonnet 4.5?
A3: Developers can access it through the Claude API using the identifier “claude-sonnet-4-5”. It is also available to regular users via the web interface and Claude apps.
