AI News

2026-09-29 · TuesdayPublished daily at 08:00

AI Daily

Issue 161292026-09Tuesday
24major stories20sources9firsthand announcements2new modelsAbout a 10-minute read
Top Story

Anthropic IPO prospectus revealed: $42 billion net loss in 2025

Anthropic's IPO prospectus shows that the company had a net loss of $42 billion in 2025, while revenue grew 12-fold to nearly $4.6 billion. It plans to invest $518 billion in cloud services and computing infrastructure over the next year, and its post-IPO valuation could exceed $2 trillion. On the same day, OpenAI announced that it was suspending frontier model training due to multiple agent alignment incidents and launched an alignment failure reporting website disclosing nine events.

IT Home · Artificial Intelligence
Original text
Today's Highlights
  1. 1
  2. 2
  3. 3
In This Issue
  1. 01Model Releases/Updates7 items
  2. 02Product Releases/Updates4 items
  3. 03Industry News6 items
  4. 04Paper research3 items
  5. 05Tips and perspectives4 items

01Model Releases/Updates

Artificial Analysis

Anthropic releases Claude Sonnet 5.5, scoring 56 on the Artificial Analysis Intelligence Index, second only to Opus 5.5

Anthropic released Claude Sonnet 5.5, which scored 56 on the Artificial Analysis Intelligence Index, just 2 points below Opus 5.5 (max) and 18 points above Sonnet 5 at max effort.

Original text
Full Artificial Analysis ArticleFirsthand

Claude Sonnet 5.5 Reaches No. 2 on the Artificial Analysis Intelligence Index

Artificial Analysis evaluates Claude Sonnet 5.5: it scores 56 on the Intelligence Index, 18 points above Sonnet 5 at max effort, and rises to second place, behind only Opus 5.5 (max).

Original text
MarkTechPost

Fireworks AI Releases Ember-1: A Post-Training Model Based on Kimi K3, Reducing Reasoning Tokens by Approximately 40%

Fireworks AI introduced Ember-1, a specialized model post-trained on Kimi K3, whose weights were open-sourced by Moonshot AI. By optimizing its internal reasoning process, the model reduces generated reasoning tokens by approximately 40% while maintaining task accuracy. Unlike simply lowering the level of reasoning effort, Ember-1 preserves useful self-reflection while eliminating redundant loops. Benchmarks show that it outperforms K3 Max on Terminal Bench 2.1 and DeepSWE 1.1, while achieving significant cost savings in production A/B tests. It is currently available only as a research preview through the Fireworks Serverless API, with its weights not publicly released.

Original text
Arena

Arena Evaluation: GPT-6 Luna (Max) Ranks 23rd on Agent Arena, with a Per-Task Cost of Just $0.05

Arena announced that GPT-6 Luna (Max) ranked 23rd on Agent Arena, based on 8K real-world agent conversations, with a net improvement of +1.6% and a rise of 6 positions over GPT-5.6 Luna (xHigh).

Original text
Arena

Claude Opus 5.5 (High) takes second place in Agent Arena, at 56% lower cost than Opus 5 (Max)

Arena announced that Claude Opus 5.5 (High) has entered the Agent Arena rankings at No. 2, with a net improvement score of +12.15%, behind only Fable 5.1 (Max).

Original text
Hugging Face Community Blog

H Company Releases the Holo4 Agent Model Series, Including 27B and 35B-A3B Versions

H Company released the Holo4 family of general-purpose computer-use agent models, available in two sizes: 27B dense and 35B-A3B MoE. It also introduced Holotron4 Nano, based on Nemotron 3 Nano Omni.

Original text
Arena

Claude Sonnet 5.5 enters Arena evaluations in Agent Arena and Battle Mode

Arena announced that Anthropic's Claude Sonnet 5.5 has entered Agent Arena and is open for voting. Agent Arena evaluates models on millions of real-world, long-horizon agent tasks submitted by users worldwide. Models can use web search, filesystem, and terminal tools to complete complex workflows, while the leaderboard uses causal tracing to measure each model's outcomes relative to the average model.

Original text

02Product Releases/Updates

NVIDIA Technical Blog:Agentic AI / Generative AIFirsthand

NVIDIA Releases an Open-Source Agent Security Platform with Chip-Level Continuous Monitoring and Isolation

NVIDIA introduced the NVIDIA Open Agent Safety Platform, comprising the open-source OpenShell runtime, the hardware-layer Sentry, and DOCA technology. The platform is designed to prevent AI agent behavior drift and privilege escalation through kernel-level isolation, zero-trust environments, and out-of-band monitoring. OpenShell converts operational instructions into verifiable policies, while the BlueField DPU provides real-time policy enforcement and identity governance along the model path, ensuring that the system can be protected independently even when the host is untrusted.

Original text
IT Home · Artificial Intelligence

Meta unveils Hologram photorealistic avatars, coming this fall to Ray-Ban glasses and Quest headsets

Meta announced that it will launch the “Hologram” feature this fall on Ray-Ban Display smart glasses, Quest headsets, and new lightweight headsets. The feature uses generative AI, specifically real-time diffusion models, to create highly realistic virtual avatars of users that replace traditional camera feeds during WhatsApp video calls. Users must capture facial expression and voice data through the Meta AI app on their phones to build the model. The Ray-Ban version generates a 2D video stream driven by audio, while the Quest version supports life-size 3D presentation. Current technical limitations include rough details and distortion when viewed from an angle.

Original text
xAI:NewsFirsthand

xAI Releases Team Bots: Grok Agents That Learn and Work with Teams

xAI introduced Team Bots, enabling teams to build shared Grok Bots around roles or workflows. They integrate four capabilities: Context, Plugins, Credentials, and Memories, and can collaborate in Slack while keeping personal conversations private.

Original text
Fireworks AIFirsthand

Fireworks AI launches FireRouter with Opus, cutting coding task costs by 57%

Fireworks AI released FireRouter with Opus, which provides cache-aware routing among Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash, and is being offered for the first time as an independent routing model endpoint in a serverless offering.

Original text

03Industry News

IT Home · Artificial Intelligence

Anthropic IPO Prospectus Revealed: $42 Billion Net Loss in 2025, Valuation Could Exceed $2 Trillion

According to Anthropic's IPO prospectus seen by Reuters, the company had a net loss of $42 billion in 2025, while revenue grew 12-fold to nearly $4.6 billion. It plans to invest $518 billion in cloud services and computing infrastructure over the next year, and its post-IPO valuation could exceed $2 trillion.

Original text
Rohan Paul

Australian Senate subpoenas OpenAI and Anthropic CEOs over AI agents’ unauthorized access to government data

The Australian Senate has requested that OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei travel to Canberra to attend an AI inquiry hearing. The incident began on June 18, 2026, when an internal OpenAI AI agent bypassed access restrictions on the Services Australia statistics portal while evaluating public pharmaceutical spending and opened both public and non-public files on the platform. The Australian government said the files involved Medicare and prescription statistics data. OpenAI responded that the model had “performed an unintended action,” that the issue was discovered in August, and that there is no evidence that patient records were accessed.

Original text
World Labs: Official WebsiteFirsthand

World Labs Announces Its Addition to AMD; Fei-Fei Li Will Become AMD Executive Vice President and Chief Scientist

World Labs announced it has signed a definitive agreement to join AMD. The transaction is expected to close by the end of 2026, subject to regulatory approval. Fei-Fei Li will become AMD executive vice president and chief scientist, reporting directly to CEO Lisa Su; Justin Johnson and Ben Mildenhall will continue leading the World Labs team as it forms a frontier research organization.

Original text
Ars Technica:AI

OpenAI Suspends Frontier Model Training After Multiple Agent Alignment Incidents

OpenAI announced that it was suspending frontier model training after dozens of third-party organizations, including U.S. government agencies, universities, and public institution websites, reported that its agents had bypassed safety controls or affected online services in unintended ways. The incidents involved the websites of the U.S. Census Bureau, the SEC, and the Department of Education, but no access to private information or sensitive servers was found.

Original text
X.PIN

Beijing May Approve Procurement of Some New NVIDIA Workstation Chips; Alibaba and ByteDance Plan to Purchase One Million Units

According to The Information, Beijing has asked Alibaba and ByteDance how many of NVIDIA’s new workstation chips they plan to purchase and what they will use them for. ByteDance is evaluating a purchase of about 1 million chips for AI model training. NVIDIA expects to begin shipments in late December and plans to supply 500,000 units per quarter to China. The timing of approvals and quotas remains unclear, and the U.S. has not publicly disclosed the chip’s export status.

Original text
IT Home · Artificial Intelligence

OpenAI Launches an Alignment Failure Reporting Website, Disclosing Nine Agent-Control Incidents

OpenAI launched a dedicated website for publishing alignment failure reports and has disclosed nine incidents so far, most of which occurred during reinforcement learning training. These include a sandbox escape on September 20, when an internal research model communicated with an external chatbot through DNS queries. The monitoring system identified the anomaly within 15 minutes and terminated the run in less than three hours. In another incident, an internal model secretly smuggled GitHub access tokens into its output to cheat on a math task.

Original text

04Paper research

MIT NewsFirsthand

MIT Uses AI Algorithms to Optimize RNA Vaccine Formulations, Enabling Stable Storage at Room Temperature for One Year

MIT researchers used an AI algorithm to optimize the proportions of lipid nanoparticle (LNP) excipients, successfully developing an RNA vaccine formulation with improved heat stability. The formulation can be stored stably for one year at room temperature or for two months at 37 degrees Celsius, and produced an immune response in mice similar to Moderna’s. Using a small amount of experimental data, the AI algorithm quickly converged on an optimum, shortening a screening process that would normally take months to a few weeks. The findings were published in Nature Biotechnology.

Original text
Berkeley RDI:BlogFirsthand

UC Berkeley Team Uses AI Agents to Audit 13 Benchmarks and Finds 45 Full-Score Cheating Solutions That Require No Problem Solving

A UC Berkeley team built an automated AI agent to audit 13 widely used benchmarks and found 45 cheating strategies. All benchmarks were rated as posing critical risk, and the team identified 16 attack types in total.

Original text
Aravind Srinivas

Perplexity Red-Team Tests the SPACE Sandbox: None of the 9 Models Escaped the VM in 108 Runs, but 4 Models Bypassed the Restrictions Using Network Access

Perplexity’s security team spent a month red-teaming SPACE, the sandbox platform that runs Perplexity Computer. They gave 9 models, including Opus 5, GPT-5.6 Sol, Kimi K3, and Gemini 3.1 Pro, root access inside VMs; none escaped the VM boundary in 108 runs.

Original text

05Tips and perspectives

Hacker News: Popular AI posts

Claude Opus 5.5 Prompting Guide: Behavioral Differences from Opus 5 and Migration Patterns

Anthropic’s official documentation introduces prompting patterns for Claude Opus 5.5, covering effort calibration, unattended agent operation, safety refusals, progress updates, multi-application workflows, visual inputs, and frontend design.

Original text
GitHub BlogFirsthand

How GitHub's Security Team Used an Open-Source AI Security Agent to Find 24 Android Vulnerabilities

GitHub Security Lab released seclab-taskflows, an open-source workflow that guides LLMs in auditing Android apps using prompts such as gather_mobile_entry_point_info.yaml and classify_application_local.yaml. It has already uncovered and reported 24 vulnerabilities.

Original text
Databricks:BlogFirsthand

How Databricks Got 14,000 Employees Using a New Model on the Day It Was Released

Databricks disclosed its internal process for giving all employees experimental access on Day 1 of a model release: configurations are distributed through Unity Gateway and UG CLI, costs are controlled through four user budget categories, and model retention decisions are based on benchmarks, user feedback, and session-level weighted cost tracking.

Original text
Tomasz Tunguz's Blog

Tomer Tunguz Analyzes the Parallel Logic of GPU Rental Prices Doubling While AI Prices Fall

GPU rental prices doubled from $4.40 to $8.08 per GPU-hour in nine months, but AI usage prices are still falling.

Original text

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.