Grok x Cursor

Over the past two years, the focus of AI model competition has been on benchmark scores and intelligence levels. However, two recent major model releases—Grok 4.5 and GPT-5.6 Sol—are sending a clear signal: users are no longer simply chasing the strongest model, but are beginning to weigh performance, cost, and reliability. This shift could reshape the entire competitive landscape of the industry.

Grok 4.5, co-trained by SpaceXAI and Cursor, is described as an Opus-level model, with performance roughly between Opus 4.7 and 4.8. But its real trump card is cost: the per-token price is 6 times lower than Opus models and 3 times lower than GPT-5.5. This means that while maintaining near-top performance, inference costs have dropped significantly. Meanwhile, OpenAI's GPT-5.6 series (Sol, Terra, Luna) is about to open to all users. Early testers report that Sol is not smarter than Anthropic's Fable, but its reliability is significantly higher—so much so that users say they 'don't want to use other models anymore.'

The improvement in cost-effectiveness is not limited to a single model. The model routing technology introduced by Factory Router has proven in weeks of production that, without sacrificing front-end performance, by automatically selecting the most suitable model for each task, it achieves a 43% aggregate cost savings, and 61% of sessions see cost reductions of at least 80%. This data, disclosed by Factory official sources, further confirms that intelligent routing can significantly optimize total cost of ownership.

The performance of open-source models is equally remarkable. In a comparative test of AI Browser Games, the open-source model Kimi K2.7 not only built games faster and at lower cost, but also had generation quality comparable to Opus 4.8. In some cases, the cost of open-source models was only 1/20 of closed-source models. This challenges the inherent belief that 'closed-source models are necessarily superior,' and means users have more cost-effective options.

Additionally, Cognition's SWE-1.7 is touted as its strongest model to date, but with lower cost and inference speed reaching 1000 tokens/s. These cases together form a trend: the AI industry is shifting from 'blindly pursuing larger models' to 'seeking the best balance between performance, cost, and reliability.'

Not all news points to a bright future. Anthropic has once again extended Fable 5 access, pushing the deadline to July 12, and claiming that after that, only prepaid credits can be used. Although this is meant to appease existing users, Ben's Bites commentary points out that this 'dangling carrot' approach actually undermines community trust in Anthropic's delivery timeline. The repeated delays of Fable 5 show that even the most advanced models can have issues with stability and commercialization pace.

This incident reminds us that a model's trustworthiness depends not only on technical metrics but also on the vendor's delivery reliability. When users face a choice between 'stronger but unreliable' and 'good enough and stable,' the latter is more likely to win long-term favor from enterprise customers. This also explains why GPT-5.6 Sol's reliability has become its biggest selling point.

Overall, cost reductions and reliability improvements are removing the last barriers to enterprise AI adoption. Grok 4.5's pricing strategy, the actual savings from model routing, and the cost advantages of open-source models will accelerate the transformation of AI from a 'showcase' to a 'productivity tool.' For enterprise users, what matters is no longer a model's position on leaderboards, but whether it can stably complete tasks within budget.

Of course, this does not mean the end of the intelligence contest. Terra and Luna from the GPT-5.6 series may bring surprises, and Fable 5 may once again take the lead after its eventual release. But the rules of the game have changed: whoever can offer 'good enough' performance combined with an attractive cost structure will take the initiative in the next round of competition.