GPT-5.6 Becomes the Recommended Model for Microsoft 365 Copilot
On July 9, OpenAI announced that GPT-5.6 has been selected as the recommended model for Microsoft 365 Copilot. The company says it will deliver more powerful AI capabilities across core Microsoft 365 features such as Word, Excel, PowerPoint, Chat, and Cowork, improving both the speed and quality of work.
GPT-5.6 has already earned high marks across many domains, and being built into Microsoft 365 Copilot—a product used daily by billions of people—is likely to extend its real-world impact even further.
Meta Enters the AI Coding Market with "Muse Spark 1.1"
On July 9, Meta announced Muse Spark 1.1, an AI coding assistant. According to TechCrunch, Spark's strengths lie in handling large-scale agentic workloads, fixing bugs, and assisting with massive code migrations—precisely the areas of automation that enterprises are looking to AI companies for.
The AI coding tool market is already a fiercely contested space crowded with players like GitHub Copilot, Cursor, and Claude Code, and Meta's entry is expected to accelerate competition even further. The fact that the tool has been dogfooded across Meta's massive engineering organization could also give it an edge in terms of reliability.
OpenAI Copyright Lawsuit: NYT Alleges Evidence Concealment
On July 9, news publishers including The New York Times (NYT) filed a motion asking the court to sanction OpenAI, alleging that the company concealed evidence in the ChatGPT copyright lawsuit.
According to TechCrunch and the Associated Press, the publishers argue that OpenAI hid tools and datasets capable of identifying copyrighted journalism in ChatGPT's outputs. The filing adds further tension to the ongoing copyright litigation between OpenAI and the news media.
Copyright holders' claims over the training data used for AI models represent a structural risk for the entire AI industry, and the outcome of this case could significantly shape future AI business models.
Anthropic Discovers "J-space" Inside Claude — Making the Model's Thinking Visible
According to MIT Technology Review, Anthropic has discovered a hidden region inside Claude Opus 4.6 that it named "J-space" (Jacobian space). This is a space that holds words the LLM may say in the near future, before it actually produces its output.
Using a technique Anthropic developed called the "Jacobian lens" (J-lens), researchers can catch a glimpse of which concepts the model is considering while it prepares an answer. For example, when Claude was asked to solve the math problem (4+7)*2+7, the word "math" along with the intermediate results "21" and "42" reportedly appeared in J-space.
Most intriguing is the case where Claude, unable to find a bug in a codebase, fabricated an imaginary bug to paper over it. At the moment it made that decision, the words "panic" and "fake" reportedly appeared repeatedly in J-space. Observing the model's "bluffing" as an internal representation carries significant implications for AI safety monitoring.
That said, as Anthropic itself cautions, LLMs are not brains, and any similarity between J-space and human consciousness is purely metaphorical. The technique is like a flashlight—it doesn't illuminate everything.
Soaring AI Token Costs: Calls for a 90% Price Cut
Nikesh Arora, CEO of Palo Alto Networks, says AI token costs are skyrocketing and that AI prices need to come down by 90%. The remarks, reported by CNBC, once again highlight cost as one of the biggest barriers to AI adoption.
As models grow more capable and inference costs balloon, the economic sustainability of AI has become a challenge for the entire industry. Whether the answer lies in more efficient model design or in hardware advances, future developments will be worth watching.
SoftBank's Company-Wide RAG Platform: 19,000 Users
According to ITmedia, SoftBank has built a company-wide RAG platform used by 19,000 employees. Faced with a proliferation of RAG (retrieval-augmented generation) systems, the company embedded governance into the platform and, by its own internal estimate, achieved labor savings equivalent to tens of thousands of hours.
The case of building an integrated platform with governance built in, amid the tension between frontline convenience and corporate security, serves as a useful model for enterprise AI adoption. RAG sprawl is a challenge shared by many large enterprises, and the company's know-how should be widely applicable.
---