September 23, 2026
Anthropic and OpenAI launch lower-priced Claude Opus 5.5, GPT-6 Sol and Luna
Officials in Pentagon inquiry link AI overreliance to Iran school strike, while ShinyHunters claims it obtained data on all FBI employees and applicants.
Anthropic and OpenAI launch lower-priced Claude Opus 5.5, GPT-6 Sol and Luna
Anthropic and OpenAI launched lower-priced AI models, sharpening competition over inference costs as businesses look to rein in spending.
- Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. GPT-6 Sol costs $2 and $10, respectively, while Luna costs $0.10 and $0.50.
- Anthropic positions Opus 5.5 as the first in a more cost-efficient Claude family. OpenAI says Sol is designed for complex work including coding, while Luna targets high-volume extraction and summarization.
- The launches follow GPT-6 Astra earlier in September and are the companies’ first model releases since Anthropic CEO Dario Amodei called for frontier development to be paced or slowed amid safety concerns.
Officials in Pentagon inquiry link AI overreliance to Iran school strike
Stale intelligence, rushed targeting and overreliance on Palantir’s AI-enabled Maven system contributed to a U.S. missile strike on a school in Minab, Iran, Bloomberg reports, citing officials involved in an unreleased Pentagon investigation.
- The site remained classified as an Iranian military facility despite operating openly as a school, while civilian-harm specialists did not review it.
- The U.S. has not publicly accepted responsibility, and Palantir denies its software was at fault.
ShinyHunters claims it obtained data on all FBI employees and applicants
Hacking group ShinyHunters claims it breached multiple FBI-related services and obtained records covering every bureau employee and applicant.
- 404 Media says it viewed a sample concerning 5,000 alleged agents, including names, home addresses, phone numbers and details about employees’ spouses.
- The breach and its claimed scope remain unconfirmed, and the report did not establish how or when the data was obtained.
Alibaba unveils Zhenwu V900 and targets over 20 gigawatts of cloud capacity
Alibaba's latest AI roadmap spans a new chip, cloud infrastructure and Qwen models.
- The company is targeting mass production and commercial release of the Zhenwu V900 in the first quarter of 2027, and says Qwen 4 is in training.
- Alibaba says it plans to expand cloud data-center capacity to more than 20 gigawatts by 2032. CNBC reported its Hong Kong shares rose around 3% after the announcements.
ByteDance tied to 73% of Nscale’s 2025 revenue in FT report
Nscale’s IPO filing disclosed heavy customer concentration without naming the customer.
- The Financial Times identified it as ByteDance, reporting that Spring, its Singapore subsidiary, agreed to rent 2,304 Nvidia B200 chips at Nscale’s Norway facility.
- The customer generated 73% of Nscale’s $33 million in 2025 revenue. Nscale warned tighter US export controls could restrict key customer relationships.
In Brief
OpenAI claims internal model solved more than 100 long-standing math problems
OpenAI says an unnamed internal model solved more than 100 long-standing mathematics problems after about a month of training. The report includes no proofs or independent assessments.
Patrick Wardle discloses zero-day in Meta’s Muse app
Security researcher Patrick Wardle disclosed a flaw in Meta’s macOS Muse app that he says lets local apps or commands expose its authentication token and hijack the privileged assistant. Meta told Ars Technica it released a hotfix.
Qualcomm announces two Snapdragon chips for on-device AI
Qualcomm announced the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6. The company says the higher-end chip can run a 30-billion-parameter mixture-of-experts model locally, though the claim has not been independently validated.
Sources say China is surveying Broadcom switches in state data centers
The Financial Times reports that China’s state asset regulator has surveyed Broadcom switch use in state-controlled data centers. One source said the review could lead to informal guidance reducing their use.
In Cisco Talos's analysis, CLOSEDQUORUM can ask AI models to choose post-compromise actions
Cisco Talos says CLOSEDQUORUM can ask up to four AI models to choose predefined actions on a compromised Windows system. Talos has not confirmed that the malware has been deployed in the wild.
British Columbia sues OpenAI over alleged failure to warn authorities about school shooter
British Columbia’s attorney-general and Tumbler Ridge’s school board sued OpenAI, the Financial Times reports. The complaint alleges the company failed to alert authorities after its systems flagged the school shooter’s ChatGPT account.
U.S. Embassy calls Australia’s proposed algorithm opt-outs censorship
The U.S. Embassy in Canberra called Australia’s draft requirement for social platforms to offer algorithm opt-outs censorship, the BBC reports. Prime Minister Anthony Albanese said the measure would return control to users.
