Anthropic just dropped Claude Sonnet 4.5, and it's making some pretty bold claims about coding endurance that sound almost too good to be true. Thirty hours of uninterrupted coding? That's longer than most human developers can stay awake, let alone productive.

The numbers behind this claim are actually impressive. Claude Sonnet 4.5 supposedly handles 64K output tokens, which translates to genuinely substantial code generation in single sessions. Compare that to its predecessor, Sonnet 4, and you're looking at 34% fewer tool calls and roughly 26% faster task completion. Not bad for a machine that doesn't need coffee breaks.

What makes this endurance possible isn't just raw processing power. The model maintains persistent state handling, meaning it remembers context and instructions throughout those marathon sessions. No more repetitive prompts every few hours. The long-context handling minimizes the constant hand-holding that plagued earlier models.

But here's the reality check: this performance comes with a price tag. We're talking $3 per million input tokens and $15 per million output tokens. Sure, prompt caching can slash costs by up to 90%, and batch processing cuts expenses by about 50%, but you're still paying premium rates compared to competitors.

The technical improvements are legitimate, though. Sonnet 4.5 outperforms GPT-5-Codex in initial reviews, demonstrating advanced reasoning and mathematical skills that actually matter for complex coding tasks. The model achieved a remarkable 77.2% score on the SWE-bench Verified benchmark, establishing itself as the best coding model available today. The model is now available alongside GPT-5 in the model picker, giving developers immediate access to compare both options directly.

The model's Context Engine uses codebase information more intelligently, and it's steerable enough to ask clarifying questions instead of charging down uncertain code paths. Multi-agent interactions add another layer of capability. The model can distribute complex tasks and maintain cooperation between AI agents over extended timeframes. These capabilities exemplify how machine learning algorithms can identify patterns and make predictions efficiently across complex development workflows.

This isn't just about one AI grinding through code for 30 hours straight – it's about sustained, collaborative development workflows. Available across multiple platforms including Augment Code, Amazon Bedrock, and Google Cloud Vertex AI, Sonnet 4.5 supports web, iOS, and Android interfaces.

The broad ecosystem support suggests Anthropic is serious about widespread adoption. Whether 30-hour coding sessions become standard practice remains to be seen. But the underlying technology suggests we're entering a new phase of AI-driven development where endurance might actually match ambition.