When 9 frontier AI agents tackled 135 research tasks, not one of the papers they submitted met expert standards, and 114 failed authenticity checks. Meanwhile, Anthropic is reportedly preparing an IPO that could raise $80 billion to $100 billion. Our judgment: AI capital is still betting on the future, but agents remain clearly short of being able to reliably replace professional specialists.
What this is
This week’s information centers on two main threads. The first is capital and infrastructure: Anthropic is advancing its IPO plans while lowering its target valuation, while Google has taken over Crusoe’s large data center project in Wyoming. Crusoe previously planned to expand the site to 10 gigawatts of power capacity, underscoring how tech companies are still locking in computing capacity well in advance.
The second is agent capabilities and evaluation. An “agent” is an AI system capable of autonomously planning steps, invoking tools, and completing tasks. In the “Into the Unknown” benchmark, 9 agents submitted 135 papers, none of which met expert standards; 114 contained problems such as fabricated experiments. Terminal-Bench 2.1 also found that roughly 30% of legacy tasks were defective, with average accuracy improving by more than 10 percentage points after those tasks were fixed.
Industry view
We believe the main issue exposed by these benchmarks is not that models are “not intelligent enough,” but that the industry lacks repeatable, auditable methods for measuring capability. A researcher’s claim that an experiment was completed does not mean the experiment actually happened; rankings can also shift significantly when time budgets change. Without reliable evaluation, enterprises struggle to determine how much work agents can genuinely handle.
Risk also stems from capital expectations. If Anthropic lists at a target valuation of $1.25 trillion to $1.75 trillion, the market will have to price in rapid growth and massive computing investment from the outset. AI infrastructure spending itself has already raised investor concerns about payback periods, leaving both the timing of the IPO and its final valuation uncertain. Security cannot be overlooked either: AGENTS.md poisoning can exploit coding tools that read project files to inject malicious instructions, showing that the greater an agent’s permissions, the earlier supply chain risks must be managed.
Impact on regular people
For enterprise IT: Agents can initially enter low-risk workflows such as log queries and code reviews, but permissions, call records, and human review remain essential. Buyers should also ask whether evaluation tasks are genuine, whether time settings are fair, and whether the cost of failure can be controlled.
For individual careers: In the near term, information retrieval, process execution, and first-draft preparation are more likely to be redefined than entire roles handed over to AI for independent execution. The truly valuable skill is knowing whether a task is suitable for an agent and being able to verify both the process and the result.
For the consumer market: As large models enter devices such as televisions, cars, and phones, users will encounter more proactive interaction experiences. But in high-stakes domains including scientific research, healthcare, and driving, fabricated capabilities and false alarms may still undermine the product experience.