Model labs pivot to agents as DeepSeek slashes V4-Pro pricing 75%
DeepSeek's permanent 75% price cut reshapes cost frontier while OpenAI and Anthropic focus on agent products.
This week's AI news underscores a strategic pivot: model labs are becoming agent labs. Greg Brockman of OpenAI signaled that the model alone is no longer the product, echoing broader industry sentiment that winning requires model + harness + workflow + UI + memory + economics. AI21 shuttered its model team to focus on agents, and DeepSeek formed a new "Harness team" for the first time. The shift means agents—AI that can take actions and use tools—are now the primary product surface.
On the pricing front, DeepSeek made its 75% discount on V4-Pro permanent, triggering market reactions. At $0.435/M input tokens and $0.87/M output, with cached input at $0.0036/M, the model now lands on the Pareto frontier for intelligence vs cost. Analysts estimate running the Intelligence Index on V4-Pro costs ~3x less than Gemini 3.1 Pro Preview, ~12x less than GPT-5.5, and ~19x less than Claude Opus 4.7. Meanwhile, OpenAI pushed Codex updates (Appshots, remote computer use, annotation mode) and Claude Devs added Sonnet 4.6 support. Gemini 3.5 Flash showed benchmark gains but mixed user feedback, with some builders noting it feels optimized for evals over human cooperation.
- DeepSeek's permanent 75% price cut on V4-Pro makes it 3-19x cheaper than frontier rivals, reshaping cost/performance.
- OpenAI's Codex update includes Appshots, /goal improvements, remote computer use while locked, and analytics.
- AI21 shuttered its model team entirely to focus on agents; DeepSeek created a new Harness team for agent integration.
Why It Matters
Agent products become the new battleground as model commoditization accelerates; cost drops enable wider deployment.