Fable 5’s Return Was Supposed to be Exciting. What Happened?

After 18 days of uncertainty and global backlash, Anthropic finally restored public access to Fable 5 on July 1 as the US government lifted export restrictions that had temporarily forced the model offline over national security concerns.

It should have marked the end of a brief but disruptive episode for developers and enterprises. Instead, it has become the beginning of a much bigger conversation.

Fable 5 returned to an industry that seems to have turned its heel. Enterprises were no longer asking only which model performed best. They were asking whether they could depend on a foreign model that could be restricted overnight. Sovereign AI, inference economics, and open-source alternatives have now become impossible to ignore.

“Relying blindly or solely on a single proprietary frontier model for mission-critical enterprise workloads is not only unnecessarily expensive but, more importantly, carries massive operational risk,” Rahul Agarwalla, Managing Partner, SenseAI Ventures, tells AIM. 

When a US regulatory directive can abruptly cut off access globally for nearly three weeks, it shatters the illusion of seamless cloud delivery. 

“When Fable was banned, a lot of enterprises started realising how risky it is to be dependent on tech over which they have no control,” says Vishnu Subramanian, Head of Product and Marketing at E2E Cloud.

Usage-Based Pricing is a Problem

Even if access remains stable, economics is emerging as an important concern. Most frontier AI providers continue to charge primarily based on token consumption. That model works for experimentation, but inference costs can become significant once applications move into production.

One catch in Fable’s return is that after July 7, it will be available only through usage credits, incurring extra costs beyond the standard monthly subscription. 

Similarly, Meta recently introduced token-based pricing for its WhatsApp Business Agent, charging enterprises based on AI usage instead of message volume, a sign that usage-based pricing is becoming central to enterprise AI. GitHub Copilot also adopted a usage-based pricing model last month.

However, the industry leaders are not happy about it.

Palantir CEO Alex Karp criticised the token-based pricing model used by AI giants OpenAI and Anthropic, saying that enterprises are paying heavily for AI services without receiving proportional value in return.

“I’m not throwing shade at them, but something has gone completely wrong,” Karp said in a recent interview with CNBC, referring to the way frontier AI labs charge customers based on token consumption. He added that many enterprise customers have adopted a mindset of wanting to “chillax and waste my time with tokens,” suggesting businesses have become overly focused on token usage instead of measurable business outcomes.

“These models have been completely, irresponsibly, oversold,” Karp said, adding that businesses are becoming frustrated with rising AI bills while struggling to justify the value generated.

Similarly, Agarwalla says that a massive amount of capital in the tech ecosystem is going towards tokens. “As enterprise use cases transition from pilot projects to massive production scales, consumption-based inference pricing becomes highly unpredictable and cost-prohibitive,” he adds.

Nikesh Arora, CEO of Palo Alto Networks, agreed that enterprises need lower token prices because high inference costs are pushing CIOs to optimise AI spending rather than explore new use cases. “If you want to win enterprise, you should be forward-pricing tokens. The cheaper the tokens for enterprises it will allow for experimentation, workflow reimagination,” he wrote on X.

That said, Roshmik Saha, Co-founder and CTO at Skyflow, a data privacy vault company, tells AIM that for many enterprises, especially early in their AI journey, consumption-based APIs remain the most efficient way to access frontier capabilities without operational overhead. However, for high-volume predictable workloads, open-source models deployed on efficient inference platforms offer a compelling alternative. 

“The future is a hybrid AI architecture. Enterprises will combine proprietary and open-source models based on the needs of each workload, balancing performance, cost, compliance, and control,” he says.

Sovereign AI is Crucial 

The Fable 5 episode, meanwhile, continues to add momentum to sovereign AI. Countries are increasingly asking whether critical AI infrastructure should depend entirely on models controlled by foreign companies and foreign governments.

Vijayakumar Arumuga Nadar, CAIO, NeevCloud, says the Fable episode is the clearest real-world validation of sovereign AI yet. “True sovereignty is not where your data lives; it is who controls the entire stack it runs on,” he says.

The debate around sovereignty is also influencing how enterprises think about AI vendors

During the CNBC interview, Karp said enterprises are increasingly questioning whether relying on companies such as Anthropic or OpenAI means giving up control over their data, intellectual property, and proprietary knowledge. He argued that businesses want ownership of their data, infrastructure, and AI models, fearing they could also hand over their ‘alpha’—their unique competitive edge.

Aggarwalla says this is a watershed moment for the sovereign AI narrative in India. He adds that for sectors such as BFSI and government, controlling where AI runs and where data resides has become a business necessity.

“If a foreign nation can pull the plug on an advanced LLM over a localised security concern, India cannot afford to depend on foreign APIs for its critical infrastructure. We will see a sharp pivot toward locally hosted AI infrastructure and localised models,” he predicts.

Can’t Ignore Open-Source Models

Fable 5’s temporary absence has also created an opening for Chinese AI companies. Z.ai (formerly Zhipu AI) recently unveiled GLM-5.2, which researchers said rivals Anthropic's Mythos in specific vulnerability-detection tasks.

“We internally started working with GLM 5.2 and OpenCode [Go]. We are making this as a deliberate effort,” says Subramanian. E2E Cloud has started working with five enterprises over the past few weeks to host open-source models and support their AI workflows.

Chinese AI companies are also capitalising on this growing demand. Companies such as DeepSeek, Alibaba, and Moonshot AI have also improved model quality while offering lower inference costs and, in many cases, deployment flexibility that appeals to enterprises.

These models are already finding their way into US companies. Airbnb has deployed Alibaba’s Qwen for customer service, Cursor has used Qwen and Kimi in its AI stack, while Coinbase CEO Brian Armstrong recently recommended Chinese models such as GLM-5.2 and Kimi 2.7 to reduce inference costs.

Sunil Gupta, Co-founder and CEO of Yotta Data Services, tells AIM that he has seen a rise in the usage of DeepSeek models on Yotta’s AI infrastructure platform, Shakti Studio. “I see a lot of adoption of DeepSeek because it is a very cost-effective and powerful model,” he says.

Reflecting this broader shift, Zoho founder Sridhar Vembu observed that the resumption of the Fable model indicates that Chinese open-source models pose a serious enough threat to market share. “I am now confident that India will catch up in AI models. There is no reason to be despondent,” he wrote on X.

Open-weight models give organisations greater control over where models run, how they are customised, and whether access can be maintained independently of a commercial API.

“Models such as Qwen, GLM, DeepSeek, and Mistral have made tremendous progress and are able to handle a wide range of production workloads, especially where fine-tuning, predictability, cost control, or deployment flexibility matter more than frontier-level reasoning,” says Saha.

He adds that frontier models can be used for complex reasoning or high-value tasks, while open models handle high-volume, well-defined workloads where control and cost efficiency matter.

The appeal of open models goes beyond economics. Nadar explains that the benefit of a model running in your own data centre is that it cannot be switched off by a letter from any foreign ministry. 

“Indian enterprises will move to a portfolio approach: frontier APIs where irreplaceable, open-weight models on sovereign infrastructure for everything availability-critical,” he says.