Anthropic Unveils Sonnet 4.5, Its Safest and Most Capable AI Model Yet
Key Points
- Anthropic launches Sonnet 4.5, touted as its safest AI model.
- Sonnet 4.5 sets a record 61.4% score in OSWorld, 17 points above Opus 4.1.
- Model outperforms Google Gemini 2.5 Pro and OpenAI GPT‑5 on coding benchmarks.
- Autonomous operation extended to over 30 hours, versus ~7 hours for Opus 4.
- Extensive safety training reduces sycophancy, deception, and power‑seeking.
- Released under AI Safety Level 3, blocking hazardous content.
- Claude Code gains checkpoint and file‑creation features.
- API pricing unchanged at $3 per million input tokens and $15 per million output tokens.
- Microsoft integrates Claude models into Copilot 365 shortly after the launch.
Anthropic announced Sonnet 4.5, positioning it as the company’s safest and most advanced AI system. The new model outperforms its predecessor Sonnet 4 and the larger Opus 4.1 on coding and agentic benchmarks, surpassing rival offerings such as Google’s Gemini 2.5 Pro and OpenAI’s GPT‑5. Safety training reduces tendencies toward sycophancy, deception, and power‑seeking, and the model now includes Level‑3 safety filters that block hazardous content. Alongside Sonnet 4.5, Anthropic refreshed its Claude Code interface with checkpoint and file‑creation features, while keeping API pricing unchanged.
Introducing Sonnet 4.5
Anthropic rolled out Sonnet 4.5, branding it as the safest AI system the company has released to date. Building on the earlier Sonnet 4, the new model is presented as the best coding model in the world, a claim backed by a suite of benchmark results. In the OSWorld suite, which measures real‑world computer tasks, Sonnet 4.5 achieved a record score of 61.4 percent, a 17‑percentage‑point lead over Opus 4.1. The model also eclipses the performance of Google’s Gemini 2.5 Pro and OpenAI’s GPT‑5 on the same tests.
Extended Autonomy and Coding Strength
Beyond raw benchmark scores, Sonnet 4.5 demonstrates a significant leap in autonomous operation. It can sustain multi‑step projects for more than 30 hours, compared with roughly seven hours for the earlier Opus 4 model at launch. This endurance is a key milestone for Anthropic’s goal of building robust agentic systems. In coding tasks, Sonnet 4.5 consistently outperforms older Anthropic models, confirming its status as the company’s top coding assistant.
Safety Enhancements
Anthropic emphasizes that Sonnet 4.5 has undergone extensive safety training. According to the company, the model is substantially less prone to sycophancy, deception, power‑seeking, and encouraging delusional thinking—behaviors that have drawn scrutiny toward competing AI systems. The new safety framework also strengthens protections against prompt‑injection attacks. Sonnet 4.5 is released under Anthropic’s AI Safety Level 3 framework, which applies filters to block outputs related to chemical, biological, and nuclear weapons.
Product Improvements Across Claude
Alongside the model launch, Anthropic refreshed its Claude product stack. Claude Code, the company’s popular coding agent, now features a new terminal interface that includes “checkpoints,” allowing users to save progress and revert to earlier states if generated code misbehaves. File‑creation capabilities, initially rolled out earlier in the month, are now directly available within chat conversations. Users who joined the waitlist for Claude for Chrome can begin using the extension immediately.
Pricing and Market Context
API pricing for Sonnet 4.5 remains unchanged at $3 per one million input tokens and $15 per one million output tokens. The announcement arrives shortly after Microsoft added Claude models to its Copilot 365 suite, underscoring Anthropic’s expanding presence in enterprise AI tools.
Looking Ahead
With Sonnet 4.5, Anthropic aims to combine heightened performance with rigorous safety safeguards, reinforcing its position in the competitive frontier‑model landscape. The company’s continued focus on developer‑friendly features and stable pricing suggests a strategy aimed at broad adoption across both enterprise and individual users.