Claude Sonnet 4.5: Coding, Long Tasks and a Simulated Desktop

I review Anthropic’s Claude Sonnet 4.5 launch, focusing on coding, longer agent tasks and a generated-interface demo. This 2025 video was sponsored by Vultr; model rankings, prices and availability are historical.

Anthropic presents Sonnet 4.5 as a leading coding model and reports results on SWE-bench Verified, Terminal-Bench, tool use, computer use and math. These are company-reported evaluations under particular setups, not proof that the model is best at every task.

I discuss METR’s work on task horizons: it measures tasks by how long a human takes and estimates how long a task a system can complete at a specified success rate. The chart in the video uses a 50% success measure on a particular task set.

I compare that chart with an Anthropic report of a model working for more than 30 hours. These are different measurements: the elapsed length of one agent run does not establish a 30-hour METR task horizon, and the run involves a surrounding agent system.

Beyond duration, I want systems to complete tasks efficiently. I discuss Greg Kamradt’s interest in action efficiency and the notion of intelligence per watt, while acknowledging that it is difficult to define one comparable measure across tasks.

I try an interactive concept demo called Imagine with Claude, which generates desktop-style interfaces from prompts. I see it as a possible direction for software, but that is a prediction rather than evidence that current applications are obsolete.

A prompt creates a futuristic email interface with simulated messages. Clicking and replying cause the demo to generate more interface content; I acknowledge that the example is a simulation and does not actually send an email.

I ask for a calculator and a to-do list, then interact with the generated views. The demo illustrates dynamically produced interfaces, although the latency and underlying implementation limit what can be inferred from a few interactions.

I ask for a browser and simulated searches. The demo generates representations of Google, Wikipedia and Reddit rather than actually browsing those sites, so its pages should not be mistaken for live web results.

I mention a reported jailbreak without reproducing the unsafe material, and show a playful website demo. A single reported jailbreak or playful output does not by itself characterize the model’s full safety or coding performance.

Box’s Aaron Levie reports testing 40,000 fields across more than 1,500 documents and finding a four-point overall improvement over Sonnet 4. This is a company-reported result for that evaluation, not an independent guarantee for other datasets.

I discuss a purportedly extracted system prompt and speculate about why it contains certain instructions. The video does not independently authenticate the full prompt, and my explanation of its authors’ motivations is conjecture.

Anthropic CEO Dario Amodei says Claude writes much of the code supporting Claude and future versions. I connect that claim to longer autonomous workflows; the video does not independently measure the proportion of code or prove an unrestricted self-improvement loop.

I read launch pricing and encourage viewers to try the model. Those figures describe the release period, and present-day costs and model availability may differ.