This has been such a busy AI news week that I did not get a chance to properly talk about Anthropic’s new Claude Sonnet 4.5 upgrade. After a day of using it (and digging into the announcement and docs; I had a lot of time in the Delta lounge and in the air yesterday), it’s clear that 4.5 is a meaningful upgrade.
Anthropic positions Sonnet 4.5 as its best model for coding, complex agents, and “computer use,” with gains in reasoning and math. Pricing stays at $3 per million input tokens and $15 per million output tokens. It’s live in the Claude API, Amazon Bedrock, and Google Vertex AI.
According to the documentation (and from what I can tell), Sonnet 4.5 runs long tasks with better state tracking, clearer progress reporting, and smarter tool use. According to Anthropic, the model can work independently for hours, coordinate parallel searches and file reads, and manage context more safely in long sessions. For heavy users, that will mean less babysitting and fewer dead ends.
Two platform features stand out. First, a Memory tool (beta) lets agents store and retrieve knowledge outside the prompt window so that projects can carry over across sessions. Second, new context-management controls automatically clear older tool calls before you hit token limits, plus a new stop reason that tells you when you ran into the context ceiling instead of a max-tokens cap. These are small switches that unlock big reliability gains.
Coding workflows got a tangible boost. Specifically, Anthropic rolled out a native VS Code extension, an upgraded terminal experience, and agent checkpoints so Claude Code can tackle longer, more complex dev work with fewer resets. Early reports and posts on X from engineers I trust call out credible improvement on coding benchmarks and multi-hour autonomy. If you have teams running agentic coding in terminals or IDEs, this is worth piloting.
The upgrade is widely available. Importantly, you can choose Sonnet 4.5 on Bedrock (AWS) and Vertex AI (GCP), which simplifies enterprise rollout if you already standardized on AWS or Google Cloud. That lowers friction for regulated shops that want managed access, logging, and standard controls.
This feels like a step-function for productivity: long-running agents that keep their place, clearer programmatic signals when you hit limits, better tool orchestration, and shipping on the clouds your teams already use. If your pilots stalled because agents forgot context, wandered, or needed constant human nudges, Sonnet 4.5 directly addresses those pain points.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.