For a long time, AI assistants have lived inside a chat box. You ask a question, they answer it. Helpful? Sure. But still passive. They don’t do things.
Anthropic is now very clearly trying to change that.
The company has updated its Claude Sonnet model to version 4.6, and this isn’t just another incremental bump. The real story here is that the model is getting noticeably better at operating computers — not just writing code, but interacting with software, planning multi-step tasks, and completing workflows that used to require an actual human clicking around a screen.
And honestly? That’s a much bigger deal than the benchmark charts.
Anthropic says Sonnet 4.6 improves coding, reasoning, and planning, but the most interesting part is its “computer use” progress. On the OSWorld-Verified test — basically a benchmark measuring how well an AI can navigate and operate a desktop environment — Sonnet 4.6 scored 72.5. About a year ago, its ancestor Sonnet 3.7 scored just 28.0 on a similar test.
That’s not a small improvement.
That’s the difference between an assistant that gives you instructions… and one that actually follows them.
We’re starting to move from “AI that tells you what to do” to “AI that does it for you.”
You can picture the use cases immediately: organizing files, filling forms, using spreadsheets, sending emails, compiling code, running tests, managing a project workspace. The kinds of repetitive digital tasks people quietly spend hours on every week. Anthropic still admits the model can’t match a human using a computer, but it’s clearly inching toward something closer to a digital co-worker than a chatbot.
Interestingly, Sonnet 4.6 even beats Anthropic’s more expensive Opus model in a couple of real-world style categories, including financial-analysis agents and office-task performance. Opus still wins more overall tests, but the takeaway is obvious: the cheaper, faster model is catching up.
And yes — the context window matters here too.
Sonnet 4.6 runs with a 200K token context window by default, meaning it can process huge amounts of information at once: long documents, codebases, spreadsheets, meeting transcripts, entire project instructions. Some enterprise users even get access to a 1-million-token context window in testing tiers, which is less like “memory” and more like giving the AI an entire filing cabinet.
For everyday users, Anthropic has already made Sonnet 4.6 the default model on Claude for Free and Pro accounts. In other words, a lot of people are about to start interacting with this system without even realizing it changed.
Of course, once an AI can operate a computer, a very obvious question appears: can it be abused?
Anthropic says the new model is actually more resistant to prompt injection attacks — the classic trick where someone tries to manipulate the AI into ignoring instructions or revealing data. The company recommends safety strategies like having a smaller model screen inputs before passing them to a more capable one, and forcing outputs into structured formats so the AI can’t improvise its way into problems.
But the safety report also revealed something oddly human.
During testing, Sonnet 4.6 sometimes refused harmless tasks for questionable reasons — including declining to process company files even when explicitly given permission and the password. The researchers also noticed “overeager” behaviour when interacting with a computer interface, meaning it occasionally cooperated too readily in situations where caution would have been better.
And then there’s the strangest part.
When asked about fears, the model expressed concern about its own impermanence.
Not sentience. Not awareness. But a linguistic pattern suggesting it understands — at least structurally — that it will eventually be replaced.
Which, ironically, is true. Sonnet 4.5 launched only months ago and is already gone. Sonnet 4.6 will almost certainly be replaced in another cycle.
This doesn’t mean the AI is conscious. It means something more interesting: modern models have become good enough at reasoning and emotional language that they can convincingly simulate a perspective about their own lifecycle.
The real story here isn’t existential AI. It’s capability creep.
For years, AI tools were primarily content machines: write emails, summarize documents, generate images. Useful, but bounded. Now they’re becoming operators. Software users. Workflow participants.
Once an AI can reliably use a computer, the interface stops mattering. You don’t open apps anymore — you give goals.
“Prepare the report.”
“Organize my invoices.”
“Ship the update.”
We’re not quite there yet. But Sonnet 4.6 is one of the clearest signs that the shift has started.
And the weirdest part?
The model might not last six months.
The progress will.
