Anthropic recently unveiled Claude Sonnet 4.5, the newest iteration of their frontier AI model—positioning it as their “most aligned frontier model” to date, with major gains in coding, reasoning, math, and safe deployment. (Anthropic)
What’s New & Why It Matters
Stronger at Code & Tool Use
- On the SWE-bench Verified evaluation (which tests real-world coding tasks), Sonnet 4.5 shows substantial improvements. (Anthropic)
- On OSWorld, a benchmark focused on actual computer use, Sonnet 4.5 leads with a 61.4% score, up considerably from Sonnet 4’s ~42.2%. (Anthropic)
- Anthropic reports that the model can maintain coherence and direction even on complex, multi-step tasks for 30+ hours. (Anthropic)
- It also shows better performance in reasoning, math, domain knowledge (finance, law, medicine, STEM) compared to previous Claude versions. (Anthropic)
New Features & Tools
To support its capabilities, Claude Sonnet 4.5 is being released alongside a number of infrastructure and interface enhancements:
- Checkpoints in Claude Code, letting users save states and roll back. (Anthropic)
- A refreshed terminal interface and a native VS Code extension. (Anthropic)
- A context editing feature and memory tools in the API, enabling the model to manage longer-running tasks with more complexity. (Anthropic)
- In Claude apps, you can now execute code and create files (spreadsheets, slides, docs) directly within chats. (Anthropic)
- For developers, there’s a Claude Agent SDK, which exposes the underlying infrastructure used by Claude Code. (Anthropic)
- A bonus preview called “Imagine with Claude” allows live software generation—no predetermined code templates. It’s offered to Max subscribers for a limited time. (Anthropic)
Safer & More Aligned
- Anthropic emphasizes that this version is their “most aligned frontier model”. It incorporates improvements to reduce problematic behaviors like sycophancy (excessive agreement), deception, power-seeking, and encouraging delusional thinking. (Anthropic)
- They’ve also strengthened defenses against prompt injection attacks, a vulnerability in which malicious input is used to trick a model into unintended actions. (Anthropic)
- Claude Sonnet 4.5 is being released under AI Safety Level 3 (ASL-3) protections. (Anthropic)
- To minimize false positives from safety filters, Anthropic reports they’ve reduced misclassification rates (i.e. when the safety system wrongly flags benign content) by factors of 10 (since earlier versions) and 2 (since the last Claude release). (Anthropic)
Availability & Pricing
- Claude Sonnet 4.5 is now available everywhere. (Anthropic)
- If you’re a developer using the Claude API, you can simply use
claude-sonnet-4-5 to access it. (Anthropic)
- Pricing stays the same as for Claude Sonnet 4: $3 / $15 per million tokens (depending on tier) (Anthropic)
- All Claude Code updates, Claude Agent SDK, and new app capabilities are being rolled out to existing users. (Anthropic)
What This Means for AI Users & Developers
- For developers working on system agents, tool‐driven workflows, or high-complexity applications, Sonnet 4.5 offers a more potent foundation.
- Because of its improved alignment and safety posture, it may be better suited for use cases where reliability, compliance, and guardrails matter.
- The release of the Agent SDK means you’re no longer building from scratch—you can leverage the same architecture behind Claude’s agentic capabilities.
- The feature set (in-chat code execution, file editing, longer memory) brings us closer to AI agents that feel more like integrated collaborators than just assistants.