Haiku 5.5 matches Chinese budget models on list price, not necessarily on the bill

Why it matters

【Key points】

Anthropic launched Claude Haiku 5.5 on Oct 8 (Beijing time) at $0.10/$0.50 per 1M input/output tokens for prompts up to 100K, matching GPT-6 Luna's short-context price. Chinese X is comparing it with domestic budget models. Editorial synthesis: official prices (read Oct 8), public posts, our judgment.

【Verifiable】

List prices per 1M tokens, in/out:

· Haiku 5.5: $0.10/$0.50 up to 100K, cache read $0.01; 5x above 100K; Batch -50%. https://platform.claude.com/docs/en/about-claude/pricing

· GPT-6 Luna: $0.10/$0.50 short, $0.20/$0.75 long context.

· DeepSeek V4.1 Flash: $0.30/$1.20 peak, half off-peak; peak cache hit $0.006.

· MiniMax-M3 (permanent 50% off, ≤512K): $0.30/$1.20.

· GLM-5.3-Flash on Z.ai: $0.15/$0.50.

DeepSeek and GLM list no long-context tier; MiniMax's starts at 512K. All three have Hugging Face weights; Haiku is API-only. Its newer tokenizer yields ~30% more tokens per text. Artificial Analysis (AA, https://x.com/ArtificialAnlys/status/2107911905822351609): Haiku at max effort scores 43 (GLM-5.3 Flash 42, Luna 38) on ~162K output tokens/task, ~3x Luna; high effort: 38 on ~55K; AA's costs omit the tiers so far.

Views: @0xLogicrw calls it a "kill line"; @mr_mofun notes the 90% cut applies only under 100K; @jacob_zhan76429 sees closed labs tiering.

【Our read】

1. Same list price ≠ same bill: token use, tokenizer and the 5x step make Haiku cheap only on short prompts at lower effort.

2. On list price MiniMax-M3 is most exposed; GLM-5.3-Flash already matches Haiku on output; DeepSeek's cache hit ($0.006) is below Haiku's $0.01, so cache-heavy, low-output agents need not pay more.

3. The Chinese edge shifts from "cheap" to open weights, self-hosting and gentler long-context pricing.

4. Test: do Chinese vendors cut prices within weeks, and how does Haiku rank once AA adds the tiers?