Model Labs Shift to Agents: DeepSeek Cuts Prices 75%

The AI landscape is undergoing a fundamental shift as all major model labs—OpenAI, Anthropic, DeepSeek, and others—are increasingly focusing on agent products rather than just base models. This was highlighted by Greg Brockman's recent comments, a reversal from previous stances. AI21 shuttered its model team to pivot to agents, and DeepSeek formed a 'Harness team' for the first time. The consensus is that model quality is no longer the moat; winning products now require a symbiosis of model, harness, workflow, UI, memory, and economics. OpenAI shipped substantial Codex updates, including appshots, /goal improvements, remote computer use while locked, annotation mode, plugin sharing, and analytics. Users report never opening an IDE in over a month, though rough edges remain with remote workflows.

On the model side, DeepSeek made a dramatic pricing move by making its 75% discount on V4 Pro permanent. First-party pricing is now $0.435/M input tokens, $0.87/M output, and $0.0036/M cached input, with a blended rate of ~$0.18/M. This places V4 Pro on the Pareto frontier for intelligence vs cost, costing ~3x less than Gemini 3.1 Pro Preview and ~12x less than GPT-5.5. Meanwhile, Gemini 3.5 Flash showed major benchmark gains but received mixed feedback from builders who found only slight real-world improvement at higher cost. The shift to agents and aggressive pricing signals that the industry is moving toward 'intelligence too cheap to meter,' fundamentally changing competitive dynamics.