Everything You Need to Know About Claude Sonnet 4.5
Anthropic released Claude Sonnet 4.5 on September 28, 2025, positioning it as “the best coding model in the world.” This update is a dramatic leap for coding capabilities, agent autonomy, and computer use that reshapes what’s possible for builders, product managers, and developers.
What Makes Sonnet 4.5 Special
Coding Excellence
Claude Sonnet 4.5 excels in coding, managing long-running agent tasks, and handling computer use. On SWE-bench Verified - the recognized standard for evaluating coding ability - Sonnet 4.5 scores 77.2% by default, reaching 82% with parallel test-time compute.
This performance sets a new benchmark that puts competitors on notice.
Focus Time
The model remains focused for over 30 hours, even during complex, multi-step projects. Developers report improvements in longer-horizon tasks, with substantial gains in planning performance and evaluation scores.
Agent Superpowers
On the OSWorld benchmark, Sonnet 4.5 scores 61.4%, a significant boost from 42.2% with Sonnet 4 just months prior.
This increase means the AI handles repetitive data entry, dashboard updates, and multi-tool workflows with minimal intervention.
In-App Superpowers
Claude now executes code and creates files - such as spreadsheets, slides, and documents - right within conversations. The workflow is streamlined: Claude produces content, we review and refine, and the task is complete.
Math and Reasoning
Sonnet 4.5 achieves 88% on AIME 2025 math challenges (up from 70.5% with Sonnet 4) and scores 83.4% on GPQA Diamond for advanced reasoning. These improvements are crucial for tasks involving financial modeling, technical analysis, and legal document review.
Claude Code Gets Checkpoints
The much-requested checkpoints feature has arrived. We can save progress at any moment, rolling back instantly if an AI-driven experiment veers off course.
Memory Tool (Beta)
A new memory tool stores and retrieves information beyond the context window, granting agents essentially limitless context. This supports knowledge base growth over time and sustains project state across sessions.
Context Editing
Sonnet 4.5 manages its context autonomously by clearing older tool calls as token limits approach. The result: no manual pruning of past interactions, with intelligent retention of essential information.
Claude Agent SDK
Anthropic has open-sourced the infrastructure behind Claude Code as the Claude Agent SDK. This tool includes solutions for agent memory management, permissions, and coordination of subagents, all vital for building production-grade agents.
What Still Needs Work
Physical reasoning remains a weak spot. Tasks demanding grounded, real-world spatial reasoning or robotics still present challenges.
Pricing: Zero Increase
Anthropic maintains its previous pricing: $3 per million input tokens and $15 per million output tokens. Despite significant performance improvements, the cost remains unchanged.
Safety and Alignment
Sonnet 4.5 is the most aligned model Anthropic has introduced, showing clear improvements against risks such as sycophancy, deception, and power-seeking behavior.
Agentic workflows benefit from greater resistance to prompt injection attacks.
Real-World Impact
Early adopters highlight transformational results:
Canva’s engineering and product teams note significant gains on complex tasks and research workflows.
GitHub Copilot observes improvements in multi-step reasoning and code comprehension for agentic experiences.
Figma reports an easier, more effective prototyping experience.
Harvey in legal tech achieves state-of-the-art performance on complex litigation, draft synthesis, and briefing analysis.
Bottom Line
Claude Sonnet 4.5 redefines expectations for AI coding assistants.
With autonomous operation exceeding 30 hours, best-in-class coding benchmarks, advanced agent abilities, and a stable price, it’s the clear choice for many developers and product teams.
Checkpoints, memory tools, and the Agent SDK equip builders with enterprise-grade infrastructure previously unavailable outside Anthropic’s own ecosystem.
👉 If you’re still using Sonnet 4 or earlier, remember to upgrade!