Anthropic on September 29 unveiled Claude Sonnet 4.5, and it’s now available on iOS. The company bills it as the “best coding model in the world,” claiming it outpaces GPT-5 Codex and Gemini 2.5 Pro on tasks involving agentic behavior, computer interaction, reasoning, and math.
What’s New in Sonnet 4.5
Anthropic didn’t just bump its version number. Sonnet 4.5 brings several features aimed at real developer workflows:
- A new terminal interface and checkpoints for Claude Code let you save progress and roll back to earlier states.
- It supports code execution and file creation (spreadsheets, slides, docs) directly in conversations.
- Anthropic unveiled a Claude Agent SDK for building custom agents.
- The model endured extensive safety training. Anthropic says it now shows reduced sycophancy, deception, power-seeking behaviors, and fewer hallucinations.
In internal tests, Sonnet 4.5 ran autonomously for 30 hours, far beyond the 7-hour limit of its predecessor.
Performance Claims vs. Market Reality
Anthropic’s pitch is quite bold: Sonnet 4.5 is superior in complex agent workflows, computer control, and reasoning over its rivals.
But where does it actually land when compared with GPT-5 and others?
- Sonnet 4.5 reportedly powers the model picker in Augment Code, offering 26% faster task completion and 34% fewer tool calls while matching Sonnet 4’s output quality.
- In some comparisons, Sonnet’s “extended thinking” mode lifts its scores above GPT-5 in specific tasks.
So far, the evidence suggests Sonnet 4.5 may outperform GPT-5 in more constrained, agentic, or interactive settings, but GPT-5 remains highly competitive in broader coding and reasoning tasks.
For iOS Users (and Developers)
- Users can access Sonnet 4.5 from their iPhones or iPads, unlocking more power on mobile.
- Developers building iOS or cross-platform tools get access to agent frameworks and code execution inline in conversations.
- With safety enhancements, Anthropic is pushing a narrative of combining raw capability with reliability.
If I were you, I’d test both Sonnet 4.5 and GPT-5 on your real-world projects. Over time, we’ll see in which domains Sonnet’s advantages truly hold.
