New GPT releases kill startups

This is the core fragility of thin AI application layers, model vendors keep moving up the stack and erase features that once looked like products. A startup that mostly wraps prompting, summarization, or document Q and A around a frontier model can lose its edge overnight when the next model follows instructions better, handles longer context, or does retrieval and tool use more natively. The durable value shifts to workflow ownership, proprietary data, and distribution.

• Jeff Tang draws the line between replaceable and durable AI companies very clearly. Fine tuning, prompt tricks, and generic customer support or copywriting bots sit closest to the model and get commoditized fastest. What tends to persist are brand, distribution, founder market fit, and data that the base model vendor does not own.
  1 sacra
• The Pinecone example shows the other side of the market. When a company sells infrastructure that many app teams need, like storing and querying embeddings, it is less exposed to any single model jump. Even there, the pressure comes from hyperscalers like AWS and Google bundling similar capabilities into their cloud platforms.
  2 sacra 4 sacra
• OpenAI itself has kept proving the point by regularly replacing older model tiers with stronger ones. GPT-4 was retired from ChatGPT in favor of GPT-4o in April 2025, and GPT-4.1 launched in April 2025 with major gains in coding, instruction following, and context length. Each upgrade shrinks the room for startups whose product is mostly model quality arbitrage.
  3 openai 5 openai 6 openai

The next wave of winners will look less like wrappers and more like systems of record or systems of action. That means products embedded in a real workflow, legal drafting, coding, sales calls, research, where usage creates proprietary data and operational lock in. As base models keep improving, the market will reward companies that own the job, not just the output.