Just hours before the deadline, Anthropic announced a last-minute reprieve: all paid subscribers’ access to Fable 5 would be extended to July 12, 2025—a five-day lifeline after weeks of uncertainty. The original cutoff, set for midnight Pacific time on July 7, had already sparked a wave of frustration among users who were planning workflows around this powerful but pricey model. But for many, including myself, this stopgap extension only reinforced a nagging question: how long can we rely on a model that feels more like a rental agreement than a tool?
Fable 5 is undeniably impressive. I used it for long-form tasks, deep research, and complex reasoning—throwing vague requirements at it and watching it churn for twenty minutes, returning results that were often more coherent and bug-free than I could have imagined. Its ability to distill messy instructions into structured outputs is unmatched. Yet its barriers are equally stark: the $10 per million input tokens and $50 per million output tokens are astronomical for any serious workload, especially when compounded by geopolitical issues like last month’s US export control suspension, which took Fable 5 offline for weeks. Access is a privilege, not a right—and for most individuals, it’s becoming an unaffordable one.
The strongest model is increasingly drifting away from ordinary users, not because it lacks capability, but because it lacks a sustainable business model for the masses.
This pattern of intermittent extensions—five days here, another five days there—creates a toxic dependency. You can’t commit your workflow to a model that might vanish tomorrow. You can’t budget for a service that keeps dangling indefinite access. It’s like being invited to a feast but never knowing when the table will be cleared.
Enter OpenSquilla’s 0.5.0 Preview. Last month, I wrote about its MetaSkill feature, which allowed agents to autonomously piece together skills into coherent workflows. The new release takes a different but complementary approach: instead of organizing skills, it organizes models.
The core idea is a multi-model integration system called Agentic Routing. Instead of betting on a single genius model like Fable 5, it assembles a team of four Chinese models—DeepSeek, GLM, Kimi, and Qwen—to tackle the same task in parallel. Each produces its own answer, then a separate model aggregates these into a final output. Think of it as an expert panel rather than a solo consultant: individually, none is the best, but collectively, they can rival a specialist—especially when the cost is drastically lower.
The project includes a technical report and a benchmark test using the open-source DRACO dataset from Perplexity—100 extremely challenging research and analysis tasks spanning finance, medicine, law, academia, and more. These aren’t simple “write me a weekly report” prompts. They require deep reasoning, fact-checking, and cross-domain synthesis. For example, one task asks to compare different econometric methods for handling heterogeneous treatment effects in staggered adoption difference-in-differences. Another asks to calculate the cash generation efficiency of CME Group based on financial data from 2024 Q1 to 2025 Q1. A third demands a differential diagnosis for a 68-year-old man with recurrent syncope and bradycardia.
Under this rigorous test, the multi-model team scored an average of 60.82 points across 100 tasks, while Fable 5 scored 59.80 points on the 94 tasks it didn’t refuse to answer (it declined 6 tasks outright). The cost difference is even more striking: the multi-model team cost $0.38 per task, while Fable 5 cost $1.21 per task—over three times as much. Not only did the ensemble match the performance of a frontier model, but it did so at roughly one-third the price, without the capricious behavior of refusing to answer certain questions.
This is not to say the multi-model approach is a perfect replacement. It introduces latency from parallel processing and aggregation, and the quality of the final output depends heavily on the averaging or voting mechanism—which can dilute strong signals when one model is clearly superior. Additionally, Chinese models may have differing performance on culturally specific or geopolitically sensitive topics compared to models trained on broader Western data. But for the vast majority of research and analysis tasks that require breadth and reliability rather than occasional brilliance, this approach is a viable, sustainable alternative.
The real breakthrough isn’t in finding a single model that can do everything—it’s in designing a system where multiple models compensate for each other’s weaknesses and amplify their collective strengths.
What does this mean for the average user? It means the era of “one size fits all” frontier models is ending. Anthropic, OpenAI, and Google are racing to build ever more powerful and expensive models, but they are simultaneously creating a ceiling on access. OpenSquilla’s approach, and others like it, are democratizing AI by leveraging what we already have—cheaper, accessible models that, when combined intelligently, can rival the best. The bottleneck, as I’ve argued before, is no longer model intelligence; it’s organization. And organization is something we can build without waiting for the next blockbuster release from Silicon Valley.
As I look at the calendar, I see July 12 approaching. Will Anthropic renew the extension again? Maybe. But I’m not betting my workflow on it. Instead, I’m shifting to a system that is predictable, affordable, and resilient—a team of models that costs a third of the price and shows up every day, no matter what geopolitics or corporate strategy decides. That’s the kind of reliability that empowers real work.