Anthropic released Claude Opus 5.5 with lower prices, faster output and vendor-reported gains in coding, computer use and long-running work. Anthropic introduced Claude Opus 5.5 on 22 September as the first model in its Claude 5.5 family. The company positions it between two familiar demands: frontier capability and a production bill that can survive sustained agent use. Its central claim is not simply that the model scores higher. Anthropic says Opus 5.5 performs at roughly the level of Fable 5.1 on most work while costing 40% less than Opus 5 on typical workloads. ...

A developer says Apple Intelligence-related software became active despite an opt-out, raising a narrow but important question about what disabling an AI feature means. Developer David Bushell published a first-person investigation on 22 September after observing Apple Intelligence-related processes and storage use on a Mac where he believed the feature was disabled. His report does not prove that every Apple device behaves the same way. It does expose an ambiguity that matters: a user-facing switch may disable visible features without removing every supporting component or background activity. ...

A new theoretical and empirical study finds two regimes where larger diffusion models can fit training data while producing worse results on unseen examples. A preprint titled “Double Descent and Malign Overfitting in Diffusion Models” examines whether familiar overfitting patterns from supervised learning also appear in generative diffusion systems. The authors report two transitions: test loss rises near a parameter-to-sample interpolation threshold and again near a parameter-to-measurement threshold. They describe the harmful behavior as malign overfitting because fitting the training objective can worsen generalization or encourage memorization. ...

A new preprint combines sparse attention with two levels of key-value sharing to reduce the memory and computation required for long-context inference. Researchers have introduced HySparse2, an attention design aimed at a practical bottleneck in long-context language models. The paper combines hybrid sparse attention with two-level sharing of the key-value cache. Its authors report experiments on an 80B-A3B model and say the method improves retrieval and downstream tasks while reducing prefill work and cache storage. ...

Meta patched a flaw that let malicious local software redirect Muse’s transcription flow and abuse the assistant’s unusually broad permissions. Security researcher Patrick Wardle found a zero-day vulnerability in Meta’s Muse assistant for macOS. According to reporting on the disclosure, another local application could alter an undocumented Muse setting and redirect speech transcription away from Meta’s service to an attacker-controlled endpoint. The attacker could then feed instructions to an assistant that already had access to the user’s apps and connected services. ...

OpenAI added two GPT-6 API tiers, giving developers a lower-cost Sol model and a very inexpensive Luna option for high-volume workloads. OpenAI’s API page now lists GPT-6 Sol and GPT-6 Luna alongside the higher-priced GPT-6 Astra. Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Both list a 1.05-million-token context window and a maximum output of 128,000 tokens. ...

Palo Alto Networks introduced a Unit 42 service that uses frontier models to find, validate and help remediate security weaknesses continuously. Palo Alto Networks has launched Continuous Frontier AI Defense, a managed Unit 42 offering that applies models from Anthropic and OpenAI to defensive security work. The company describes a loop that discovers weaknesses, validates whether they are exploitable, proposes remediation and tests again. It is an attempt to make offensive-style testing persistent instead of a project scheduled once or twice a year. ...

Bloomberg reports that the Pentagon linked overreliance on AI-generated analysis to a missile strike on an Iranian school that killed civilians. A Bloomberg investigation published on 22 September says the US Defense Department acknowledged that excessive reliance on AI-generated analysis contributed to a deadly strike on a school in Iran. If the account is accurate, it is a rare official connection between an AI-supported targeting process and a mass-casualty error. The underlying public finding was not available during this review, so the central claim remains verified as Bloomberg reporting rather than independently confirmed fact. ...

Snorkel AI raised $350 million at a $3.5 billion valuation as companies spend more on specialized data for training and evaluating AI systems. Snorkel AI has raised $350 million in a financing round that values the data-development company at $3.5 billion, according to Reuters. The report says annualized revenue has risen above $350 million from about $20 million a year earlier. Those figures describe a rapid expansion in demand for the less visible work around models: building, labeling, filtering and evaluating data for particular domains. ...

Unreal Labs released an open-source agent harness that runs tool calls asynchronously and claims lower costs than Codex on selected benchmarks. Unreal Labs has published Unreal Agent, a coding and general-purpose agent harness designed to issue independent tool calls concurrently. The company argues that many agent loops wait unnecessarily for one command to finish before starting another. By identifying operations that do not depend on each other, the harness can reduce idle time and sometimes total token use. ...

Online AI discussion on 22 September centered on the cost of generated text, uncertain measurements of model behavior, local image generation and dramatic media framing of safety incidents. » Why it matters: Community forums often surface useful experiments and criticism before formal publication. They also mix evidence, demonstrations and opinion, so each item below is labeled by what it can support. Hacker News Attention becomes the scarce resource The most active thread linked an essay arguing that abundant media makes human attention more valuable. Commenters ranged from browser-history design to the way language models lower the cost of producing text. This is worth attention because it frames an economic problem for publishing: readers pay the review cost even when an author pays almost no production cost. ...

Reuters reports that Aikido Security released an open-weight model for code security that organizations can run locally instead of sending repositories to an external provider. Aikido Security, a Belgian application-security company, is moving part of AI-assisted code analysis closer to the customer. The reported model is designed for cybersecurity tasks and local deployment. The central promise is data control: sensitive source code can remain inside the user’s environment while the model examines it. ...