Huawei has unveiled a new test for artificial intelligence agents called Claw-Anything, and the results are sobering. The benchmark simulates months of digital life for an AI assistant, demanding it handle long-term tasks across that extended period. GPT-5.5, described by the company as the best model currently available, scored just 34.5%.
What the benchmark measures
Claw-Anything isn't a single-round quiz. It places an AI agent in a simulated digital environment and runs it through a series of tasks that unfold over what amounts to months of virtual time. The agent has to keep track of context, remember earlier interactions, and adapt as the scenario evolves. Huawei says the design mirrors the kind of persistent, evolving existence a real digital assistant would need to manage — not just answering isolated questions but sustaining coherent behavior over a long stretch.
The benchmark's creators argue that most current tests for AI agents are too short. They measure quick reasoning or pattern matching but don't capture whether a model can maintain a thread across dozens or hundreds of exchanges. Claw-Anything aims to fill that gap by stretching the timeline and piling on complexity.
GPT-5.5's performance
GPT-5.5, which Huawei calls the best model it tested, reached 34.5% on the benchmark. That number means the model successfully completed just over a third of the tasks it was given. The rest it either failed or did only partially. The company hasn't released scores for other models, so there's no direct comparison, but the single figure makes one thing clear: even the top performer has a long way to go before it can handle sustained, long-term digital work reliably.
The low score isn't necessarily a surprise. Long-term memory and context management remain hard problems for large language models, even as they ace short-form tests. GPT-5.5 can write essays and answer complex questions, but keeping a coherent thread over simulated months is a different kind of challenge.
Why the number matters
Benchmarks like Claw-Anything shift the conversation away from flashy demos and toward real-world utility. An AI that forgets what it did last week isn't much use as a persistent assistant. The 34.5% score suggests that even the best current models lose their grip when the timeline stretches.
Huawei's test doesn't solve the problem, but it puts a number on it. That number — 34.5% — gives developers a target to beat. It also raises a practical question for anyone building products around AI agents: how much can you trust a model that fails two out of three long-term tasks?
The company hasn't said when it will open Claw-Anything to outside researchers or whether it plans to update the benchmark with new scenarios. For now, the results stand as a reminder that the path to truly persistent AI is still a steep climb.
