Community Trust ScoreVerified
Huawei built a benchmark. It’s called Claw-Anything. And the results are pretty rough for the AI industry.
The test dropped AI assistants into simulated digital environments — basically fake but detailed versions of the kind of digital life a person manages every day. Scheduling, decision-making, task execution, context-switching. The kind of stuff humans do without thinking twice. GPT-5.5, currently the most advanced AI model available, cleared only 34.5% of the benchmark’s tasks. Not a typo. Thirty-four point five percent. For a model that’s supposed to represent the cutting edge of what AI can do right now, that’s a hard number to spin.
Not great.
What Claw-Anything Actually Tests
Huawei designed Claw-Anything specifically to stress-test AI assistants on digital life management. The benchmark doesn’t just throw logic puzzles at a model or ask it to write code. It simulates the kind of messy, context-dependent decision-making that real digital existence demands — the sort of thing where the right answer depends on who you are, what you were doing five steps ago, and what you’re probably trying to accomplish next.
That’s a different animal from most AI benchmarks. A lot of standard tests reward raw reasoning or pattern-matching. Claw-Anything seems to care more about adaptability. Can the model handle a situation it hasn’t seen cleanly before? Can it manage competing priorities without losing the thread? Can it behave, basically, like a person navigating a normal digital day?
GPT-5.5 mostly couldn’t. Or at least, it could only about a third of the time.
The gap between what the benchmark demands and what the model delivered is wide. Huawei’s design essentially asks: if we handed you someone’s digital life to manage, how often would you get it right? The answer, for the best model currently out there, is less than four times in ten.
That’s the headline.
Why a 34.5% Rate Is a Big Deal
There’s a temptation to read a benchmark result and shrug. Benchmarks get gamed. Tests get criticized. Numbers get reframed. But 34.5% is hard to reframe as anything other than a ceiling problem for current AI systems.
It’s not that GPT-5.5 is a bad model. By most standard measures, it’s the best one available right now. The point is that “best available” still falls well short of what would be needed to genuinely manage a digital existence the way a human does. The model can probably handle isolated tasks well enough. It’s the integration — the sustained, adaptive, contextually-aware management of a full digital environment — where things fall apart.
And that gap matters for anyone thinking seriously about AI agents, AI assistants, or the broader idea of handing AI meaningful autonomy over digital tasks. The Claw-Anything results are a reality check. Maybe a necessary one.
Huawei didn’t release an immediate comment on next steps after the benchmark results came out. No roadmap, no follow-up timeline. Unclear whether further iterations of the test are planned or whether the results will feed directly into model development guidance.
Where AI Development Goes From Here
The AI industry has spent a lot of time lately talking about agents — models that don’t just answer questions but actually do things on your behalf. Book the meeting. File the document. Manage the inbox. Handle the workflow. The pitch is compelling. The Claw-Anything results are a reminder that the pitch and the reality are still pretty far apart.
For AI to work as a genuine digital life manager, it probably needs to get a lot better at a few specific things. Contextual memory — keeping track of what happened earlier and why it matters now. Adaptive prioritization — figuring out what matters most when tasks compete. And something harder to name but easy to recognize, the ability to handle ambiguity without defaulting to a wrong answer confidently.
GPT-5.5 at 34.5% means the current generation of models hasn’t cracked those problems. Not even close, really.
Benchmarks like Claw-Anything are useful precisely because they’re hard to game. When a test simulates a full digital environment rather than a narrow skill, models can’t just pattern-match their way to a high score. They have to actually perform. And performance, right now, is limited.
The broader AI development community will probably pay attention to these results. Huawei’s benchmark is a specific kind of pressure test, and a 34.5% ceiling on the top model is the kind of data point that shapes where research money and engineering effort go next.
Whether the next generation of models does meaningfully better on Claw-Anything — unclear. No one’s said. But the bar is set now, and it’s sitting at 34.5%.
Frequently Asked Questions
What is the Claw-Anything benchmark?
Claw-Anything is a benchmark designed by Huawei that tests AI assistants by placing them in simulated digital environments, evaluating their ability to manage tasks and decisions the way a person would in a real digital context.
How did GPT-5.5 perform on the Claw-Anything benchmark?
GPT-5.5, currently the most advanced AI model available, achieved a success rate of 34.5% on the Claw-Anything benchmark, the highest score among tested models.
