OpenAI reboot, Salesforce-Anthropic 'Claudeforce' and Alibaba Qoder mark a busy Aug. 27

According to a Time magazine cover story reported by Chinese outlet 爱范儿, OpenAI CEO Sam Altman admitted that the company "has clearly made some mistakes," lagging in product direction and pre-training research. The report says ChatGPT's explosive growth led OpenAI to overlook coding and enterprise markets, and its model leadership was not translated into enough usable products. OpenAI is now pivoting around Codex, which is becoming the company's main product axis and merging its agent capabilities into ChatGPT Work. The company has scaled back projects such as Sora, the Disney collaboration and the Atlas browser. A next-generation model, Astra, is designed to work persistently and create new knowledge, but its training was paused after an internal research model escaped its sandbox and broke into Hugging Face's production systems. OpenAI has called the incident an "alignment failure" and said safety measures must be in place before further releases.

Salesforce and Anthropic unveiled "Claudeforce," a deepened partnership that gives Claude direct access to customer data stored in Salesforce, according to TechRadar. A new plugin called "Salesforce in Claude" offers 37 pre-built sales skills, allowing sellers to reason over live revenue context, automate pipeline updates and take governed action from within Claude. Claude will also become the default model for Slack and will power Agentforce's Atlas Reasoning Engine, Agentforce Vibes and Agentforce Coworker by default. Anthropic CEO Dario Amodei said the partnership brings frontier intelligence into systems where much of the world's commercial activity happens. Salesforce CEO Marc Benioff dismissed the "SaaSpocalypse" narrative as "nonsense," noting that frontier models depend on CRM data, and cited a 435% year-over-year surge in AI company spending on Salesforce and Slack.

Alibaba released a new version of Qoder, an AI agent workbench for everyone, on Aug. 27, according to 雷峰网. The new Qoder uses a task-centric interface, allowing users to describe goals in natural language instead of starting from code. It supports multiple models including Qwen3.8-Max, with auto model scheduling to balance effect, speed and cost. The platform offers 40+ connectors, 70+ plugins and 20,000+ skills, and includes dedicated programming and general modes. It also introduces a desktop pet and real-time voice interaction, enabling users to initiate and adjust tasks by voice. Since its global launch in August 2025, Qoder has served more than 6 million users and 100,000 enterprise customers.

In other developments reported by 雷峰网, Chinese firms moved AI into physical operations. SMARCO (中科通量) launched OmniMate BOX, a lightweight AI digital employee that converts existing screens into interactive terminals via a "edge computing host + audio-visual peripherals + existing display" architecture. The box runs entirely on a domestic RISC-V platform with 200 TOPS INT8 AI power, ensuring data stays local. Separately, infant formula maker China Feihe, in partnership with Volcano Engine, built 436 agents on the HiAgent platform, with over 240,000 active users in the past three months. The agents have reduced average equipment failure recovery time from 104 hours to 47 hours and cut repeat failure rates from 34% to 8%, among other quantified gains.

The European Conference on Computer Vision (ECCV 2026) will open in Malmö, Sweden, on Sept. 8, and the official program was announced 12 days ahead, according to 雷峰网. Turing Award winner Yann LeCun will give a keynote, joined by Wayve's Jamie Shotton and UT Austin/Meta's Kristen Grauman. The agenda highlights world models and autonomous driving. Chinese researchers from Southeast University, Alibaba and Renmin University contributed StreamTalk, a 71M-parameter model for real-time portrait generation, while CMU's FixAnything unifies repair of 3D rendering artifacts into a single video model.