The gist: Major AI labs like OpenAI, DeepSeek and others are pivoting from pure model development to integrated agent systems. The combination of model, harness, workflow and UI becomes the actual product. At the same time, coding agents are improving noticeably, while DeepSeek’s aggressive pricing puts the market under pressure.
The industry is undergoing a fundamental transformation: leading AI labs are no longer developing just models, but increasingly integrating agent systems as a core product. This signals a strategic shift toward complete AI products rather than isolated models.
The leading AI development labs are realigning their strategies. While OpenAI works toward its forthcoming IPO filing, influential voices like Greg are showing a remarkable course change: model labs are increasingly building agents as a core product. This represents a clear break with the long-dominant philosophy of major model developers.
This development manifests concretely in several steps: AI21 has shut down its model division and is focusing entirely on agent systems. Even DeepSeek, the renowned Chinese lab, is building its own “harness team” for the first time. The trend is unmistakable – systems are triumphing over pure models.
The new reality defines successful products through a symbiosis: model quality alone no longer creates a competitive advantage. Winners are products that seamlessly integrate model, harness (training framework), workflow, user interface, storage and business model. An increasing number of experts emphasize that the model alone is no longer the product.
The development carries a subtle danger: if models become so specialized through proprietary harnesses that they only function effectively with closed agent systems, vendors can increasingly steer users toward their agents – at the expense of open API competition.
In coding agents, concrete progress is emerging. OpenAI delivered significant updates with “Codex Thursday No. 6”: app screenshots, improved /goal functions, remote access on locked systems, annotation mode and extended analytics. Users report significant workflow disruptions – some have not opened their IDEs for weeks. Claude is meanwhile expanding auto mode to the pro plan and supporting Sonnet 4.6.
On the pricing side, DeepSeek is sending massive signals. The permanently anchored 75% discount on DeepSeek-V4-Pro fundamentally changes the cost boundary. With estimated $0.18 per million mixed tokens, V4 Pro sits on the efficiency frontier and costs roughly three times less than Gemini 3.1 Pro Preview, twelve times less than GPT-5.5 and nineteen times less than Claude Opus 4.7. The market is responding to a world in which intelligent AI becomes too cheap to count.