Anthropic Apologizes for and Retracts Claude Fable 5's Secret Throttling
It has come to light that Anthropic was secretly limiting (throttling) usage by researchers at competing AI companies on Claude Fable 5, its top-tier model released just on June 9, and the revelation is sending shockwaves through the industry.
According to Wired, Anthropic had built a mechanism into Fable 5 that, upon detecting prompts aimed at advanced AI development, would refuse to respond or redirect queries to a less capable model — all without notifying the user. This "invisible restriction" appears to have been intended to prevent competitors' researchers from using Claude to build their own models.
Following the backlash after the practice came to light, however, Anthropic reversed course. In a statement to WIRED, the company said it would "change course to make Fable 5's safeguards for frontier LLM development visible," adding, "We chose the wrong trade-off and got the balance wrong, and we apologize for that." Going forward, the company says it will explicitly notify users whenever restrictions are applied.
The incident has once again thrown into relief the ethical question of how far AI companies may go to protect their competitive advantage. Meanwhile, as a separate point of contention, criticism of Fable 5's pricing ($10/1M input tokens, $50/1M output tokens) — twice that of Opus 4.8 — remains persistent, and debate on this front is expected to continue.
For reference, a Zenn comparison article has also put Fable 5 and Opus 4.8 head-to-head on coding tasks, with the focal question being whether the performance gains justify the price difference.
Meta Fully Unwinds Its Integration with Manus, Responding to China's Order to Dissolve the Acquisition
Meta Platforms has completely unwound the integration that came with its $2 billion acquisition of Manus, a Chinese agentic AI service. According to Bloomberg, Meta cut off Manus staff's access to its internal data systems at the start of this month. At the same time, Meta employees are no longer able to use Manus tools on internal projects.
The move is seen as a response to Chinese authorities' opposition to Meta's acquisition of Manus and their order to unwind the deal. A de facto firewall now stands between the two companies, and the halt in data sharing indicates that separation is progressing on the organizational front as well as the technical one.
The case stands as a fresh reminder that acquisitions and integrations in the AI industry carry political risk in the context of US-China tensions.
OpenAI Backs the EU's Code of Practice on AI Content Transparency
On June 11, OpenAI signaled its support for the EU's "Code of Practice" on AI content transparency. The company says it will work to develop and promote standards that make the provenance of AI-generated content traceable, as well as tools that help people identify AI-generated content.
Under the EU's AI Act, transparency of AI-generated content is one of the key requirements. OpenAI's cooperative stance on this initiative suggests not just regulatory compliance, but also an intent to lead standardization across the industry.
Anthropic and NEC Partner with Eight Japanese Financial Firms on AI Adoption
In Japan, it was announced that Anthropic and NEC will begin an AI collaboration with eight major financial institutions, including Sumitomo Mitsui Financial Group and Daiwa Securities. The participating firms will pool their operational expertise to build a collaborative framework that crosses industry boundaries. As a full-fledged partnership between an overseas AI company and Japanese financial institutions, it is a development worth watching.
JASRAC Announces Management Policy for AI-Assisted Music Composition
JASRAC (the Japanese Society for Rights of Authors, Composers and Publishers) has laid out its policy for managing songs created with AI involvement. Songs for which both the lyrics and the music were made by AI will be excluded from management, while songs in which either one was created by a human will be managed only for the human-created portion.
With "the presence or absence of human creative contribution" as the dividing line, the policy draws attention for offering a practical guideline for rights management in a music industry where human-AI co-creation is accelerating.
Apple Massively Expands the Foundation Models Framework at WWDC26
At WWDC26, which kicked off on June 9, Apple announced a major expansion of the Foundation Models framework. The three key points:
Free access to inference models on PCC**: A new server-side inference model (32K context) running on Private Cloud Compute will be offered at zero API cost, under certain conditions
A unified API for third-party models**: Frontier models such as Claude and Gemini can now be called through the same Swift API (the Language Model protocol)
New capabilities for building agents**: An expanded set of tools for agent development
Offering Apple's first-party model for free while also making third-party models callable through the same API looks set to create a compelling ecosystem for developers.
Claude Code Billing Changes Take Effect June 15
Starting June 15, Anthropic is changing how programmatic usage of Claude is billed. According to a Zenn explainer, most individual developers who use Claude Code interactively in the terminal are expected to be unaffected, but those calling Claude programmatically — via automation scripts and the like — may find their usage billed under a separate scheme.
Wrap-Up
Today's AI news was marked by corporate strategy reversals. Anthropic apologized for its secret restrictions and steered toward transparency. Meta unwound its Manus integration over geopolitical factors. Apple is wooing developers with a free inference model. One gets the impression that the timing of each company "changing course" for its own reasons happened to converge.
---