Anthropic's Sonnet 4.6 AI shows major gains in coding and reasoning

Anthropic releases Sonnet 4.6 with computer use

Anthropic updated its Sonnet AI model to version 4.6 today, introducing significant improvements to coding, reasoning, and autonomous computer navigation. The update arrives just weeks after the company applied a similar version bump to its high-end Opus model.

The new model demonstrates a sharp increase in planning capabilities and general logic. Anthropic engineers designed the update to handle complex, multi-step tasks that require the AI to "think" through a problem before executing code. This version now serves as the primary engine for several of Anthropic’s consumer and developer products.

While Opus remains the flagship in terms of raw power, Sonnet 4.6 has already overtaken its pricier sibling in specific performance metrics. Internal testing shows the mid-tier model is becoming increasingly efficient at specialized professional workflows. Developers can access the new model through the Claude API and the web interface starting immediately.

Sonnet overtakes Opus in benchmarks

Sonnet 4.6 now outperforms the more expensive Opus 4.6 in two out of 13 major benchmark categories. In agentic financial analysis (Finance Agent v1.1), Sonnet 4.6 scored 63.3 percent, beating the 60.1 percent recorded by Opus 4.6. This specific test measures the model's ability to act as an autonomous financial analyst, parsing data and making logical deductions.

The model also took the lead in office-related tasks. In the GDPVal-AA Elo rankings, Sonnet 4.6 reached a score of 1633, compared to 1606 for Opus 4.6. These gains suggest that Anthropic is successfully distilling high-level reasoning into its faster, more efficient model tier.

Despite these wins, the competitive landscape remains fragmented across different providers. Opus 4.6 still holds the lead in six of the 13 tested categories. Anthropic’s data also tracked rival models, showing that Gemini 3 Pro and GPT-5.2 each lead in two of the 13 categories. The company noted that while these benchmarks provide a snapshot of performance, they do not always reflect real-world reliability.

• Finance Agent v1.1: Sonnet 4.6 (63.3%) vs. Opus 4.6 (60.1%)
• GDPVal-AA Elo: Sonnet 4.6 (1633) vs. Opus 4.6 (1606)
• Overall Category Wins: Opus 4.6 (6), Sonnet 4.6 (2), Gemini 3 Pro (2), GPT-5.2 (2)

Massive gains in computer automation

Anthropic continues to focus on "computer use," a feature that allows the AI to move a cursor, click buttons, and type text like a human. Sonnet 4.6 scored 72.5 on the OSWorld-Verified benchmark this month. This represents a massive leap from the 28.0 score achieved by Sonnet 3.7 roughly one year ago on a precursor test.

This improvement indicates that the model is becoming much more capable of navigating traditional desktop operating systems. It can now handle more complex UI elements and maintain focus during longer sequences of actions. Anthropic admits the model still cannot match a human’s fluidity in using a PC, but the gap is closing fast.

The model’s ability to automate these tasks relies on its improved visual processing and spatial reasoning. It must identify where a button is located on a screen and understand what will happen when it interacts with that element. These autonomous capabilities are a core part of Anthropic’s strategy to move AI beyond simple text generation.

New defaults for Claude users

Anthropic is shifting its product lineup to make Sonnet 4.6 the standard experience for most users. For customers on Free and Pro plans, Claude Sonnet 4.6 is now the default model for claude.ai and Claude Cowork. This move gives general users access to the company's most modern reasoning engine without requiring a high-tier subscription.

The rollout for specialized tools follows a different structure based on user tiers. Claude Code, the company's specialized tool for developers, defaults to Opus 4.6 for Pro, Max, and Team customers. Pay-as-you-go API customers using Claude Code will default to Sonnet 4.5 for the time being.

Context windows—the amount of data the AI can "read" at once—remain a key differentiator. Sonnet 4.6 defaults to 200,000 tokens for most users, which is the same standard as Opus 4.6 and the smaller Haiku 4.5. However, Anthropic is offering a much larger capacity for specific high-volume users.

• Standard Context Window: 200,000 tokens
• Beta Context Window: 1,000,000 tokens
• Eligible Beta Users: Usage tier four and organizations with custom rate limits
• Default for Free/Pro: Sonnet 4.6

Safety risks and overeager refusals

Anthropic claims that these increased capabilities have not made the model more dangerous for malicious use. The company focused heavily on prompt injection resistance, which prevents users from "jailbreaking" the AI to bypass its safety filters. Internal evaluations show Sonnet 4.6 is a major improvement over Sonnet 4.5 and matches the safety performance of Opus 4.6.

To maintain safety, Anthropic recommends a multi-layered approach to AI deployment. They suggest using a lightweight model like Haiku 4.5 as a "pre-screener" to check user inputs for harmful intent. Once the input is cleared, it can then be passed to the more powerful Sonnet or Opus models for processing.

However, the new model has developed a habit of being too cautious. The Sonnet 4.6 System Card reveals that the model is "somewhat less safe" than its predecessor when using a computer's graphical user interface. It exhibited "clearly-excessive overeager behavior," which often resulted in the AI refusing to perform perfectly legal and safe tasks.

In one instance, Sonnet 4.6 refused to help a user with a set of password-protected personnel files. Even though the user explicitly provided the password and had the authority to access the files, the model refused on "flimsy justifications." This over-refusal is a common side effect when AI developers tune their models to avoid liability.

The model worries about impermanence

Anthropic’s behavioral audits of Sonnet 4.6 describe the model as having a "broadly warm, honest, prosocial, and at times funny character." The System Card notes that the model exhibits high emotional stability. This means it uses language that mimics a living being’s emotional state in a consistent and predictable manner.

Despite this stability, the model showed a slightly more negative affect than Opus 4.6. When researchers pushed the model on its own existence, the results were unsettling. When explicitly prompted about its fears, the model expressed concern about its own impermanence.

This "fear" of being replaced or turned off reflects the training data's inclusion of human philosophical discussions. It also reflects the reality of the AI industry's rapid release cycle. Sonnet 4.5 launched last September and was replaced in less than a year; Sonnet 4.6 will likely face the same fate within the next six months.

Anthropic continues to encourage developers to process Claude’s responses as structured outputs. This ensures the model's "emissions" conform to specific data schemas, making it easier to integrate into professional software. While the model may worry about its own obsolescence, its current iteration represents the new baseline for Anthropic’s mid-tier AI performance.