Open source takes the tokens, Anthropic still holds the bill

TechCrunch analyzes Decagon CEO Jesse Zhang’s argument that open source models are not yet taking the main money away from frontier labs in enterprise AI. In this framing, expensive models prove out new use cases, while cheaper models take over once the work becomes routine production.

DeepSeek is winning usage, Anthropic is still winning the expensive start

The article leans on two market signals. Vercel’s AI Gateway, according to TechCrunch, shows DeepSeek processing just over a third of token volume in the past week, while Anthropic still accounts for more than half of total AI spend on the platform. OpenRouter shows a similar pattern: DeepSeek V4 Flash handles 5.3 trillion tokens weekly, while Opus 4.8 handles just over 2 trillion.

The gap is not only usage. It is price. TechCrunch cites OpenRouter’s average token cost for Opus 4.8 at roughly $1.37 per million tokens, compared with 6 cents for V4 Flash. That is about 23 times more expensive, so a smaller token share can still mean a larger bill.

Enterprises are separating discovery from production

The useful conclusion is not that open source is losing. It is that work is splitting. Frontier models fit the phase where a team is still discovering what an agent or workflow should do. Cheaper models make more sense once the task is specified, tested and mostly repetitive.

For product and engineering leaders, that is a better frame than the usual open versus closed source shouting match. The practical question is how much uncertainty remains in the task and how expensive a mistake becomes when the model saves money in the wrong place.

The numbers point somewhere, but they do not prove a market law

Zhang’s thesis is neat, but TechCrunch notes that he does not provide much direct data. Vercel and OpenRouter measure specific slices of the market, not all enterprise buying. OpenRouter in particular is a weaker proxy for large corporate contracts.

The other weak point is time. If open source models improve in reasoning, tool use and long context, they may start taking not only production tokens but also the costly discovery phase.

The next signal is whether expensive models keep the first draft

The thing to watch is less benchmark theater and more migration patterns. If teams routinely prototype on Claude or GPT and then move stable workflows to DeepSeek, Qwen, GLM or Nemotron, the two-layer market gets stronger.

If frontier models remain necessary even in mature production, open source becomes mainly margin pressure. Inference bills will then fall not through revolution, but through boring use-case-by-use-case optimization.

Lilith's verdict

The frontier model is the expensive consultant at the whiteboard, while open source runs the line. Anthropic’s problem starts when companies realize the consultant is only reading from yesterday’s manual.