Alex Good: 40 Times Cheaper Qwen 3.6 is Redefining AI and Open Source Asset Pricing
Crypto investor and AI practitioner Alex Good (goodalexander) stated on social media that his cost for using Anthropic Claude Opus is about "$80 per terminal per hour, totaling $160 per hour for two terminals," while the cost of the new model based on Alibaba's Qwen 3.6 is "around $2 per hour," making it approximately 40 times cheaper, yet it is "close to or even reaches the current strongest level" in multiple tasks. He emphasized that the community is seriously undervaluing the pricing of open models and low-cost cutting-edge capabilities, especially for crypto AI projects that require large-scale self-deployment and decentralized operation, where closed-source models are nearly infeasible in terms of technology and compliance; open-source or open-weight models are the truly viable infrastructure.
Public benchmarks and community tests show that Alibaba's Qwen 3.6 series (especially Qwen 3.6 Plus) is close to or even surpasses closed-source flagship models like Claude Opus in various metrics in scenarios such as agentic coding and complex task execution, providing a combination of "performance close to, but price far below" in benchmarks like terminal automation, SWE-bench, and Terminal-Bench 2.0. This model of "cutting-edge performance × affordable or even free" is increasingly seen by developers as a practical production alternative. Multiple technical evaluations and blogs point out that Qwen 3.6 Plus has significant cost advantages in throughput, latency, and long context capabilities, making it feasible to achieve a combination of "running close to Opus experience on a laptop or personal GPU," further strengthening the voice of open-source and low-cost models in the global AI landscape.
Source: Public Information
ABAB AI Insight
Alex Good's core point is not that "a certain Chinese model is strong," but rather the cost structure: when a model with "capabilities close to Opus" is priced at 1/40, this is not just a product difference but is reshaping the economic foundation of global AI computing power and models. For crypto AI scenarios that require continuous running of agents, automated trading, on-chain risk control, and quantitative research, hourly cost differences can amplify into order-of-magnitude capital threshold differences over weeks/months—whoever can replace a "$160/hour" model with a "$2/hour" model can run more experiments, train more agents, and iterate faster within the same budget, creating speed and scale effects.
This directly challenges the pricing logic that "closed-source models = the only cutting-edge." In the past, the market was willing to pay a premium for Opus-level models due to clear advantages in complex reasoning, reliability, and tool usage; however, the Qwen 3.6 series has narrowed or even leveled the gap in agent benchmarks, terminal tasks, and long context scenarios, leading users to start viewing "closed-source cutting-edge" as an option that is "slightly better but extremely expensive," rather than an absolute necessity. Once this perception spreads among developers and startups, the premium for the "strongest models" will be reduced to only a small portion of scenarios where it has a decisive impact on task results, while the majority of applications will naturally shift to cheaper open-source/open-weight models.
From the perspective of crypto and decentralization, his mention that "we cannot support closed-source models in the long term" is a structural constraint rather than an emotional judgment. Decentralized networks that want to embed AI at the protocol layer must be able to self-deploy models in off-chain nodes, user devices, or decentralized computing markets, which requires models to be controllable in terms of weight licensing, operating costs, and compliance risks; the closed-source API model is inherently concentrated in the hands of a few large American companies, which contradicts the crypto logic of "anyone can replicate execution on any node." Models like Qwen 3.6, which have controllable costs and can run on personal hardware, provide a more aligned technical path for "crypto x AI": the model itself is embedded in the protocol, rather than a centralized API's debt.
On a deeper level, this is also a realistic aspect of the "competition between Eastern and Western AI technologies and capital": leading Western companies maintain high prices and high profit structures under closed-source and security narratives, while Chinese manufacturers rapidly occupy developers' minds and actual usage volumes through more aggressive open-source/open-weight strategies and extremely low marginal prices. If this situation of "40 times price difference, limited performance gap" continues to exist, the global AI computing power and application layer may resemble the pattern of the Android and white-label phone era: high-end closed-source models occupy brand and profit, while open-source and low-cost models capture the vast majority of real traffic and usage time, with ecosystems like crypto, which are extremely cost-sensitive and naturally reject centralized dependencies, almost purely betting on the latter.