Anthropic Cuts Fable 5 Costs Through Delegation, China Clears Limited Nvidia H200 Purchases, and Coinbase Halves Costs with 1,200 Agents — AI News Digest

Anthropic's "Advisor pattern": Using Fable 5 as a manager to rein in costs

Claude Fable 5, Anthropic's most capable model, took first place in every category of the industry-by-industry performance index (six sectors including finance, law, and healthcare) published by Artificial Analysis. That performance comes at a steep price, though. On the Strategy & Ops Index, it costs $3.48 per task — a gap of more than 100x compared to DeepSeek V4 Pro's $0.03.

To address this cost problem, Anthropic recommends an "Advisor pattern" in which Fable 5 serves as the planner and delegates actual execution to the cheaper Sonnet 5. According to the company, this configuration delivers 92% of Fable 5's standalone performance at 63% of the cost. Rather than applying the expensive model to every task, the approach dedicates it to planning, keeping production costs at a realistic level.

The guidance signals that the price-to-performance ratio of frontier models has become a key differentiator in real-world deployments.

China clears limited purchases of Nvidia H200 chips

According to Bloomberg and The Information, the Chinese government has informed major AI companies — including Alibaba, ByteDance, and DeepSeek — that it will allow limited purchases of NVIDIA's H200 chips. Companies must apply for approval, specifying how many chips they need and what they will be used for.

The H200 is a high-performance processor for AI model development, and U.S. export controls have restricted its supply to Chinese companies. The easing can be read as a move to balance sustaining China's domestic AI development against its dependence on U.S. technology. Purchase volumes will remain limited, however, and government control through the application and approval process is expected to continue.

Coinbase runs 1,200 AI agents and halves costs with a multi-model strategy

Cryptocurrency exchange Coinbase has reportedly halved its AI costs through a multi-model infrastructure strategy while running 1,200 AI agents. Rather than depending on a single high-performance model, the company built a system that dynamically selects and switches to the right model based on the nature of each task.

At a scale of more than 1,000 agents, optimizing model selection has an enormous impact on cost. Coinbase's case shows that operating AI agents in the enterprise is gradually shifting from "which model to use" to "how to combine multiple models."

Prime Intellect raises $130 million Series A to let companies build their own agents

Prime Intellect, founded in 2024, has raised $130 million in a Series A round. The company's goal is to provide a platform that lets enterprises build and train their own agent systems without relying on frontier AI labs.

Demand for building AI agents in-house is rising, as Coinbase's example confirms. Prime Intellect's approach could be an attractive option for companies that want to avoid the risk of being locked into platforms like OpenAI's or Anthropic's. The size of the round reflects how much growth the market expects.

NVIDIA's Nemotron 3 Ultra posts the top accuracy for an open model on a LangChain benchmark

NVIDIA announced that Nemotron 3 Ultra has achieved the highest accuracy of any open model on a Deep Agents harness tuned by LangChain. It reportedly matches leading closed models on business-task accuracy while costing less than one-tenth as much for inference.

What's notable is that these results were achieved without retraining the model. The fact that harness optimization alone brought performance close to that of closed models suggests that the maturity of inference infrastructure and agent design has become a critical factor in model selection.

HuggingFace releases a natively integrated vLLM backend for transformers

HuggingFace has announced an effort to integrate a vLLM-based modeling backend into the transformers library, enabling inference at native speed. Developers will be able to take advantage of vLLM's fast inference engine without significantly changing their existing transformers code.

In the local LLM community, the choice of inference engine directly affects performance, so integrating vLLM with transformers could substantially improve the developer experience. A detailed technical write-up is available on the HuggingFace blog.