Last week, OpenAI launched a model it calls “the most intelligent and aligned” it has ever built. A Chinese startup unveiled the first AI model designed from the ground up to use tools, not just talk about them. And a telecom operator proved its AI model had been downloaded five million times.
The common thread? AI is no longer being judged by how well it chats. It’s being judged by what it can actually do.
GPT-6 Astra: OpenAI’s “Most Intelligent” Model Yet
On September 3, OpenAI officially launched GPT-6 Astra, calling it the most intelligent and aligned model the company has ever developed. The benchmark numbers are striking: on Agent’s Last Exam, which measures an AI’s ability to complete complex, multi-step professional tasks, Astra scored 59.3%, surpassing the previous flagship GPT-5.6 Sol.
But the more telling metric came from NVIDIA CEO Jensen Huang, who shared a test where GPT-6 Astra, after 60+ hours of operation, scored 14 out of 100 on a 3D spatial reasoning benchmark. It was the first AI model to achieve a double-digit score—and Huang commented that AGI has effectively arrived.
The model is designed for complex reasoning and computer use. It can fill out forms, update CRM records, test software, and analyze scientific data. This isn’t a chatbot upgrade. It’s a worker upgrade.
NeoHorse-1: The First Agent-Native AI Model
On September 8, a startup called TokenRhythm—founded by the former head of Huawei’s Noah’s Ark Lab—released what the industry is calling the first Agent-Native model. NeoHorse-1 comes in 4B and 9B versions, and it’s built on a fundamentally different principle.
Traditional AI models learn from textbooks and Q&A data. NeoHorse-1 learns from real agent task execution traces—when to search, when to call a tool, what the tool returns, what to do next, and how to recover from mistakes.
The approach works. In benchmark tests covering agent harness, tool use, code, and instruction following, the 4B version ranked first among comparable models, surpassing the Qwen3.5-9B base model on five benchmarks. The 9B version showed particular strength in complex state maintenance, long-range debugging, and failure recovery.
The training infrastructure was provided by InfraEngine (无问芯穹), which optimized inference costs by 90% over the past year. The message is clear: AI is learning to work, not just to answer.
DeepSeek V4.1-Flash: The Smallest Model Ahead of a Big IPO
On September 10, DeepSeek launched V4.1-Flash, the smallest model in its new architecture family, designed for faster inference, higher throughput, and scaling to larger models. The release came as the company prepares for an initial public offering on Shanghai’s STAR Market.
The timing is strategic. DeepSeek has been one of the most closely watched Chinese AI labs since its R1 model disrupted industry assumptions about cost. By releasing a smaller, efficient model ahead of its IPO, the company is signaling that it can deliver capability without requiring massive compute.
OTel 2.0: The Telecom AI Model Downloaded 5 Million Times
For years, telecom operators struggled with AI models built for general purposes, not for networks. AT&T’s OTel 2.0 changed that.
Released in partnership with Dell and AMD, OTel 2.0 is the largest and best-performing open-source model built specifically for telecoms. It understands telco language, 3GPP standards, and the complexity of modern networks—built from the ground up rather than adapted after the fact.
Since release, it has been downloaded more than 5 million times, building on OTel 1.0 models that were downloaded nearly 30 million times. AT&T processed more than 1 trillion tokens to generate roughly 440 billion training tokens.
This is what open telco AI looks like at scale. Not a pilot. A production-ready model built for the realities of the network.
Nokia and Microsoft: Agentic AI for Network Automation
On September 17, Nokia extended its partnership with Microsoft to build an agentic, unified data foundation for telecom operators. The solution integrates Nokia Data Suite with Microsoft Fabric, allowing operators to access high-quality, trusted data in minutes instead of weeks.
Initial use cases focus on RAN optimization through autonomous VoNR assurance and geo-experience, correlating subscriber, network, and RF data to pinpoint coverage or capacity issues. The collaboration enables intelligent, agent-based solutions across the network stack.
“Together, we are helping networks evolve from static infrastructures into programmable, AI-native platforms,” said Vivek Jaiswal, Senior Vice President at Nokia.
Valdyr: The AI-Native Operating Layer for Telcos
On September 15, Swedish company Telness Tech rebranded to Valdyr and repositioned its Seamless OS as the “native AI execution layer” for telecom operators.
The distinction is important. Most telecom software records and reports decisions. Seamless OS executes them. Offers, subscriptions, billing, service provisioning, customer journeys, and system integration all run on a single platform. Telecom business processes are built as automated workflows rather than custom development projects.
The company is already working with Telia and Truecaller, and its pitch is simple: AI-native operations shouldn’t require a three-year reconstruction project.
Accenture and Google: 1,000 Engineers for Agentic AI
On September 8, Accenture and Google Cloud launched the Accenture Gemini Enterprise Business Group, committing 1,000 forward-deployed engineers to help enterprises scale agentic AI. The group is backed by Accenture’s nearly 50,000 Google Cloud-skilled professionals.
Real-world results are already emerging. During NFL Sunday Ticket demand, a Gemini Enterprise customer-service agent increased customer sentiment by 11% and cut average handle time by 37%.
This is enterprise AI deployment at scale. Not pilots. Production systems handling real traffic.
Baidu’s “Super Agent Family” and the 2.2 Billion Agent Prediction
On September 9, Baidu Cloud unveiled its “Super Agent Family” at a conference in Suzhou, positioning agents as the unified human-machine collaboration entry point for enterprises. The company cited a prediction that by 2030, more than 2.2 billion agents will be in use globally, with 81% of Chinese enterprises already deploying agents into core production management.
One example: a Suzhou consumer electronics manufacturer used Baidu’s agent to solve warehouse management challenges, raising direct-loading fulfillment rates from under 60% to over 80% and increasing daily shipment speed by 30%.
Six stories from one week. A “most intelligent” model from OpenAI. The first Agent-Native model from China. A telecom AI downloaded 5 million times. Agentic network automation from Nokia and Microsoft. An AI-native execution layer for telcos. And 1,000 engineers deployed to scale enterprise agents.
#AI, #OpenAI, #GPT6Astra, #NeoHorse, #DeepSeek, #AgenticAI, #TelecomAI, #Nokia, #Microsoft, #AT&T, #GoogleCloud, #Accenture, #Baidu, #AIInfrastructure, #DigitalTransformation, #Innovation, #TechNews