LLM API price wars have reached an absurd point.

DeepSeek V4-Pro's cache-hit input price dropped to 0.025 RMB per million tokens -- by comparison, GPT-5.5's equivalent input price is roughly 0.88 RMB. That means DeepSeek's price has slashed to 1/35th of GPT. On the other side, Anthropic's Claude Fable 5 didn't cut prices but raised them -- input $10/output $50, prices effectively doubled, pursuing a "most expensive but best" premium route.

One slams into the floor; the other shoots for the ceiling. Model vendors in the middle are getting squeezed from both sides.

Two Ways to Survive

An Altimeter Capital partner put it bluntly: the LLM market ultimately leaves only two survival modes -- "either be outstanding, or be very cheap."

The outstanding ones, like flagship Claude and GPT, earn premium pricing from top users through cutting-edge reasoning, strongest code generation, most reliable long-context understanding. The data speaks volumes: Claude Code's top 10% power users contribute 80-90% of revenue. These users don't care if the unit price is $5 or $50; they care about "getting the job done right."

The very cheap ones, like DeepSeek and various open-source fine-tuned models, carry massive token consumption through extreme cost control. 80% of everyday tasks -- text classification, simple summarization, routine customer service -- don't need flagship IQ; a fast, cheap model is enough.

Model Routing Becomes the Core Pattern

This stratification is spawning a critical new layer: the Model Router.

Legal AI unicorn Harvey has already provided a template: through a hybrid routing scheme -- assigning simple tasks to open-source fine-tunes and routing complex tasks to Opus 4.7/4.8 -- they exceeded pure Opus results on some scenarios while slashing costs. This isn't a "one model rules all" approach; it's "let the right model do the right job."

In future AI application architectures, model selection itself becomes a core competency. You don't need to bind to one model vendor; you need a smart "dispatcher" -- it knows what model to use for what task, knows when to save money and when to spend it.

What It Means for Developers

For regular developers and founders, pricing polarization brings two clear takeaways:

First, don't waste time building a model that's "a bit cheaper than GPT but about as good." That position no longer exists. Either you can achieve DeepSeek-level extreme cost, or Claude-level extreme capability -- otherwise it's a dead end.

Second, the model routing layer is an underappreciated opportunity. When models become utilities like water and electricity, "helping you pick the right model" itself becomes commercially valuable. Just as the cloud computing era spawned cost-optimization companies like CloudHealth, the AI era will spawn new players in model routing and cost optimization.

Price wars will continue. But the ending is coming into view: LLMs won't be a "winner takes all" market, but a stratified one -- the top tier sells IQ, the bottom tier sells compute, and the middle sells routing and services.

See you tomorrow.