For two years the “price butcher” of large models was never Claude. DeepSeek drove per-million-token prices to cents and trained Chinese vendors to match. On 8 October the positions flipped. Anthropic launched Claude Haiku 5.5, a lightweight model for everyday tasks, at USD 0.10 in and USD 0.50 out per million tokens for requests under 100,000 tokens, one tenth of the prior generation and cheaper than DeepSeek-V4.1-Flash.
On the same day OpenAI said GPT-6 was fully live, with paying users on GPT-6 Sol and free users on GPT-6 Luna. Luna also lists at USD 0.10 in and USD 0.50 out, but OpenAI pushed it into the free ChatGPT tier, so consumers pay nothing. One vendor dropped to the floor, the other went free, and both Western giants moved on price at once.
The fight is over small models, the entry point for developers. On API bare price for short context, Haiku 5.5 and Luna sit below DeepSeek-V4.1-Flash, though DeepSeek’s cache price of USD 0.003 still beats Haiku’s USD 0.01 for heavy document reuse. Benchmarks are vendor-reported and not directly comparable, but practitioners found the three small models close in ability, with differences only on complex instructions.
Why give it away? Anthropic treats the entry tier as a break-even traffic engine, not a loss leader, then earns back through dearer tiers: Haiku jumps to five times output past 100,000 tokens, Opus 5.5 input is 40 times the Haiku floor, and Sonnet 5.5 cache-read just dropped from USD 0.20 to USD 0.10 to keep coding-agent bills down. OpenAI’s free Luna is capped and small-model-cheap, so the compute cost stays low while the user lands in its funnel.
The layered model is now consensus among Western labs. Google moved first, with Gemini 3.1 Pro at USD 2 in and USD 12 out and a free Flash-Lite tier. The risk falls on Chinese vendors, whose paid penetration is low and whose main revenue is API and ToB. DeepSeek’s API stack is the most complete, but it lacks a consumer paid story, while Qwen, Kimi, Doubao and Zhipu each lean on open weights, private deployment or traffic.
The lesson is structural. Western labs can price a model at USD 0.10 because subscriptions, cache, tools and enterprise sit behind the gate. The Chinese vendors that win will be the ones that build the toll road, not just the cheaper token.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1085815599_116132.