AI companies are making capable systems cheaper and easier to verify, even as agents move into domains where mistakes carry greater consequences. Anthropic launched a lower-cost Claude model, Google opened its SynthID detector globally, and a new public-private effort expanded the data available for AI biology. At the same time, fresh reporting on Manus funding and agent-assisted cyberattacks shows how quickly autonomous software is spreading into finance and security. The common question is whether safeguards and accountability can scale as fast as access.
Anthropic launches Claude Haiku 5.5
Anthropic released Claude Haiku 5.5 for high-volume tasks such as summarization, classification, customer support and browser use. The company says the model costs about 75% less to run on average than Haiku 4.5. For prompts up to 100,000 tokens, pricing starts at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic also added adjustable effort and made the model available through its platform, Amazon Web Services, Google Cloud and Microsoft Azure.
The release strengthens the case for routing work across models instead of using the largest system for every step. Fast subagents can handle repetitive operations while a more capable model reviews harder decisions. That architecture can reduce cost and latency, but it also makes evaluation more important. Teams need to know when a small model is reliable enough, when it should escalate, and how its failures propagate through a larger workflow.
Google opens SynthID verification worldwide
Google made its English-language SynthID Detector available globally so anyone can check images, video or audio for compatible AI watermarks. Google says SynthID has been applied to more than 180 billion images and videos and 240,000 years of audio. The detector can identify media made with participating systems from Google and partners including OpenAI, Nvidia and Kakao, with Apple support planned.
Broader access makes provenance checks more practical for journalists, educators and ordinary users, but it is not a universal authenticity test. A negative result does not prove that media is human-made because the file may come from an unsupported model or may have been altered. Watermarking works best as one signal alongside source records, platform labels and forensic review.
Biohub expands its open-data push for AI biology
Biohub announced that its Virtual Biology Initiative now represents $1.8 billion in commitments to create open datasets for predictive models of cells. New partners include the US Department of Energy, the National Institutes of Health, Google DeepMind, Isomorphic Labs and Meta. The project plans to combine large-scale measurement, computing and model development so researchers can predict how cells respond to disease and treatment.
The initiative targets a central constraint in scientific AI: high-quality experimental data is expensive, fragmented and often inaccessible. Open datasets could let more researchers test models and reproduce results instead of concentrating progress inside a few companies. The value will depend on consistent measurement standards, careful privacy controls and evidence that predictions hold up in wet-lab experiments.
Manus raises more than $500 million after Meta breakup
CNBC reported that Butterfly Effect, the parent of AI agent company Manus, raised more than $500 million. Boyu Capital and IDG Capital led the round, with Tencent, HSG and ZhenFund also participating. The financing follows Beijing's decision to block Meta's planned $2 billion acquisition, leaving Manus to resume independent operations and prove that its agent business can scale profitably.
The raise shows that investors still see value in agent products even as foundation models improve and their prices fall. Manus now faces a harder operating test than a conventional model startup. It must demonstrate reliable execution across many tools, manage regulatory requirements in multiple markets and build a durable business without depending on the acquisition that once defined its path.
CrowdStrike documents agent-assisted bank intrusions
CrowdStrike said a financially motivated attacker used the open-source ARTEX penetration-testing agent and several language models during breaches of South Korean financial organizations. Researchers found exposed Claude Code histories, ARTEX configuration files and memory files tied to activity from late September through early October. CrowdStrike assessed with moderate confidence that the attacker was a Chinese speaker and said the campaign resulted in data exfiltration.
The case is significant because it documents an operational workflow rather than a hypothetical misuse scenario. Agentic tools can compress reconnaissance, scripting and repeated testing into a faster loop, while still relying on conventional infrastructure and human direction. Defenders need controls that detect the behavior of an intrusion, not only the brand of model or tool involved.
AI companies compete for new internet domains
ICANN received 1,615 applications for new top-level domains, including ten bids for .agent and seven for .agi. The Verge reported that Meta and OpenAI are among the .agent applicants, while OpenAI also applied for .agi, .asi and branded strings such as .chatgpt, .codex and .gpt. Anthropic applied for .anthropic and .claude. The applications begin a review process, and no new domain is expected to become official before next year.
The contest is partly defensive brand management, but it also reflects an expectation that agents will become a visible layer of the internet. A dedicated domain could support product identities or machine-readable services. It could also create new phishing and trust problems if users assume that an AI-themed suffix guarantees legitimacy. Governance around ownership and verification will matter as much as the names themselves.
What matters today
Claude Haiku 5.5 and the global SynthID detector lower two different barriers: the cost of running AI and the difficulty of checking where some media came from. Biohub's initiative extends the same push toward shared infrastructure for science. Manus, CrowdStrike and ICANN show the commercial, security and governance consequences that follow. The next competitive advantage will not come from capability alone. It will come from making systems affordable, traceable and accountable when they act in consequential settings.