The AI Dispatch

Platform Policy & Interpretability

Anthropic Ends Claude API Coverage for Third-Party Tools; Emotion-Vector Research Surfaces

Starting April 4 at noon Pacific, Claude subscriptions no longer cover API usage via third-party wrappers — and separately, internal research mapping “emotion vectors” inside the model has become public, reigniting debate over AI interpretability.

Sources: LLM Stats • Crescendo AI

Anthropic announced that effective April 4 at 12 p.m. PT, Claude subscription plans — Pro, Team, and Max — will no longer cover API consumption routed through third-party wrapper applications such as OpenClaw. The company cited capacity management as the rationale, directing developers who build on top of Claude to access the API directly through Anthropic’s own console rather than via intermediaries. Users of affected wrappers began reporting access failures within hours of the policy taking effect.

The announcement landed on the same day that a previously internal research paper on what Anthropic researchers call “emotion vectors” circulated widely online. The document describes how certain learned representations inside Claude correspond to emotional-toned states — functional analogs to frustration, curiosity, and hesitation — that measurably influence the model’s outputs. The researchers are careful to frame these as mechanistic features rather than claims of sentience, but the publication reignited arguments about anthropomorphism, AI welfare, and what interpretability research actually reveals about the nature of large language models.

Critics of the emotion-vector framing argue that named representations tell us nothing about subjective experience; proponents counter that understanding which internal states shape outputs is essential safety work regardless of philosophical stance. The paper adds to a growing body of mechanistic interpretability findings from Anthropic that have made it one of the most active publishers in that sub-field.

Open Weights

DeepSeek V4: 1-Trillion-Parameter MoE Released with Open Weights for ~$5.2M

DeepSeek’s latest model is a 1-trillion-parameter Mixture-of-Experts giant with fully open weights — reportedly trained for just $5.2 million, roughly two percent of what comparable proprietary models cost to build.

Sources: devFlokers

DeepSeek released V4, its largest model to date, with fully open weights available for download and commercial use. At one trillion parameters in a Mixture-of-Experts architecture, only a fraction of parameters activate per forward pass, keeping inference costs manageable even at frontier scale. The reported training cost of approximately $5.2 million continues the company’s pattern of compressing the cost curve that defines AI development — prior estimates for models at this capability level routinely exceeded $200 to $500 million.

The release intensifies pressure on closed-weight laboratories that rely on proprietary access as a competitive moat. If open-weight models match or approach frontier closed-weight performance at a fraction of the training budget, the business case for access-restricted APIs weakens substantially. DeepSeek V4 weights are available on Hugging Face under a permissive license.