Meta's Internal AI Costs Reach Billions, Shifting to Token-Level Budget Management

An internal memo sent by Meta CTO Andrew Bosworth to 6,000 employees has revealed that the company's internal AI usage costs have reached into the billions. In the memo, Bosworth stated that "not all movement is progress, and token usage alone is not a measure of impact," signaling a plan to reassess how AI is used across the company.

Specifically, starting in 2027, the company plans to set token consumption budget caps for each department and introduce a system for centrally managing usage through a central dashboard called "AI Gateway." While Meta has actively encouraged internal AI use until now, cost-side governance is only just beginning.

This move shows that AI running costs are becoming a serious challenge even inside major tech companies. Alongside techniques for reducing inference costs, organizational token management looks set to become a key theme going forward.

German Court Finds Google Liable for AI Overviews

A German court has ruled that Google itself bears legal responsibility for false statements generated by Google's AI Overviews (the AI-powered search result summary feature). The ruling stated that "companies that design, train, operate, and control AI systems bear legal liability for damages arising from the responses those systems generate."

This ruling could become an important precedent for where legal responsibility lies for AI-generated content. Platform operators have traditionally taken the position that they are "mere intermediaries of information" with limited liability, but when it comes to responses that AI itself generates, the trend toward holding designers and operators accountable is strengthening.

As the phased rollout of the AI Act progresses in Europe, the scope of operator liability for AI outputs is also being defined concretely in the courts. Attention will likely focus on similar rulings in other countries and the implications for Japan.

Claude Fable 5 Scores 88% on FrontierMath's Hardest Tier, Beating GPT-5.5 by 13 Points

It has been reported that Anthropic's Claude Fable 5 achieved an 88% accuracy rate on the hardest tier of FrontierMath, a math benchmark. This is a dramatic improvement over the under-10% figure that Opus 4.5 recorded in early 2026, showing that the pace of improvement in AI mathematical ability is accelerating.

Meanwhile, OpenAI's GPT-5.5 scores about 75% on the same tier, giving Fable 5 a 13-point lead. That said, superiority on a single benchmark does not directly translate into performance on real-world tasks, and comprehensive evaluation across cost, latency, safety, and other factors remains essential.

Even so, the pace at which accuracy on the hardest math problems jumped from under 10% to 88% in less than half a year is astonishing. Among researchers, some are now saying that "the ceiling of AI capability in mathematics has become impossible to see."

GLM 5.2 Deployed with 1M Context and 2 Thinking Modes (Update from Previous Post)

[Updated from our previous coverage of the GLM-5.2 MIT license release]

It has been revealed that z.ai's GLM 5.2 has been deployed on the GLM Coding Plan, now supporting a 1 million-token context window and two thinking modes (max / high). The max mode is recommended for coding use cases. The API and MIT-licensed open weights are reportedly scheduled to be released within a week.

Third-party benchmarks have already started coming in, and where GLM 5.2 will land in coding performance among open models is drawing attention.

---