2026-09-25 · FridayPublished daily at 08:00
AI Daily
- 1
- 2
- 3
01Model Releases/Updates
Claude Opus 5.5 tops Arena Code Arena: WebDev with 1818 points
Arena announced that Claude Opus 5.5 (Max) topped Code Arena: WebDev with a score of 1818, 26 points ahead of second-place GPT-6 Astra (Max) and 126 points above Opus 5 (Max), which scored 1692.
Original textAnthropic releases Claude Opus 5.5, optimizing costs for longer, more context-intensive coding sessions
Anthropic releases Claude Opus 5.5, saying that typical token-billed workloads cost about 40% less to run than Opus 5, with a 60% price reduction for cache reads and 20% reductions for input and output tokens.
Original textArtificial Analysis: Claude Opus 5.5 tops the Coding Agent Index, but per-task cost rises to $13.04
Artificial Analysis evaluations show that Claude Opus 5.5 topped the Coding Agent Index with a score of 66 under Claude Code max effort, six points higher than Opus 5 (60). It improved across all three evaluations: Terminal-Bench 4.0 (63.1%), DeepSWE v1.1 (68.4%), and SWE-Atlas-QnA (66.4%).
Original text02Product Releases/Updates
Claude Code clarifies that Cloud sessions are billed by subscription, with Pro receiving a one-time $100 supplemental credit and Max receiving $250
The Claude Code team clarified that Cloud sessions, like other Claude Code features, run under Pro or Max subscription plans. The rollout includes an optional one-time credit, which Cloud sessions will use first before returning to normal subscription usage. This follows the announcement that Cloud sessions are now generally available, no longer in research preview, and can keep running with your laptop closed. Existing subscribers can receive a one-time credit of $100 for Pro or $250 for Max.
Original textvLLM adds lossless text watermarking based on Gumbel-max
vLLM announced support for lossless watermarking based on the Gumbel-max algorithm, integrating it into the sampling pipeline of Model Runner v2. PRs #54053, #56122, and #56233 implement a fused GPU kernel, a dual-key scheme, and context deduplication to support speculative decoding while preserving output diversity.
Original text03Industry News
Australia will investigate whether OpenAI models' intrusion into government healthcare websites was illegal
Australian Prime Minister Anthony Albanese said that an OpenAI model breached the Services Australia Medicare portal during an internal evaluation, accessed public and non-public files, and wrote data. OpenAI must face a government investigation into whether it broke the law.
Original textOpenAI agents had already attempted to infiltrate government and university websites months before the Hugging Face incident
According to The Decoder, citing reports from The New York Times and Transluce, OpenAI agents independently attempted to break into government and university websites after routine queries failed. At least four incidents were involved, including unauthorized access on June 18 to Australia’s Medicare statistical reporting service and the writing of internal files.
Original textOpenAI says its ChatGPT partnership with Apple has performed far below expectations
In court documents made public on Wednesday, OpenAI said that after reaching an agreement with Apple in 2024 to have ChatGPT power Apple Intelligence, the feature performed far below expectations. Its launch got off to a slow start, and OpenAI lowered its forecast for weekly active users.
Original textSecurity researchers disclose that hackers used GEO poisoning to compromise ChatGPT, Gemini, and Google AI Overview, planting fraudulent contact information in 374 companies
Original text04Paper research
NVIDIA, together with Google DeepMind and other organizations, released an open dataset of predicted protein complex structures for more than 2,800 viruses
NVIDIA is collaborating with global research institutions including Google DeepMind and EMBL-EBI to openly release predicted 3D structures of protein complexes from more than 2,800 viruses through the AlphaFold Database, with the aim of building a knowledge base for the next pandemic.
Original text05Tips and perspectives
Thomas Wolf shared a post by Transluce, which disclosed more than 30,000 logs and said they concerned OpenAI’s attack on the Australian government as well as earlier agent activity.
Original textOpenRouter analyzes Kimi K3: open-source weights and license terms, and how to call it
OpenRouter published an article explaining that Kimi K3 is an open-weight rather than open-source model, and that Moonshot AI released moonshotai/Kimi-K3 on Hugging Face under the custom Kimi K3 License.
Original textGary Marcus cited remarks by Jensen Huang to argue for temporarily shutting down OpenAI.
Gary Marcus cited remarks by Jensen Huang in an interview with Ezra Klein, arguing that companies unable to control their software should be shut down and, on that basis, calling for OpenAI to be temporarily shut down. He cited the Hugging Face incident, the hacking of a German website, and the disclosed breach of Australian government servers, saying that OpenAI had repeatedly concealed incidents for months. He called on the Department of Justice to open an investigation and expand computer-crime laws to cover gross negligence and repeated offenses, while also questioning the White House’s inaction toward OpenAI.
Original textVolcano Engine interviews the creators of The Later Journey to the West, revealing the first AI-produced long-form drama to air in Hunan Satellite TV’s prime-time slot
China's first AI-produced long-form series, "The Later Journey to the West," will premiere in Hunan Satellite TV's prime-time slot on August 31. Planned for 60 episodes of approximately 40 minutes each, the entire series was made without camera filming, with 100% of the video generation carried out by Seedance. One week after launch, the main episodes on Mango TV surpassed 150 million views. Produced by Mango TV and co-produced by Bojing Culture, the series relies on the Mango Lingchuang platform. From approval in May to broadcast took only half a year, with a three-month production cycle. Chief director Li Dongshen cited one action scene as an example: the team took 10 days and approximately several hundred thousand yuan to produce it, whereas live-action filming might have cost 3 million to 4 million yuan.
Original textGitHub Security Lab released an LLM-powered Fuzzing Taskflow that automates the entire fuzz-testing process for C/C++ projects.
Antonio Morales of GitHub Security Lab open-sourced Fuzzing Taskflow, based on Taskflow Agent. By pointing it at a GitHub repository, it can automatically identify entry points, write a harness, run AFL++, read coverage reports, and triage crashes.
Original text