Anthropic's Haiku 5.5 prices at $0.10 per million input tokens — matching GPT-6 Luna at the floor
Claude Haiku 5.5 lands as Anthropic's smallest, cheapest model yet — and its arrival at the exact same input price as OpenAI's rumored Luna tier signals a deliberate floor being drawn across the industry.
On October 7, 2026, Anthropic released Claude Haiku 5.5 — and the number that matters most isn't the model's benchmark score. It's $0.10 per million input tokens.
That's the same reported input price as GPT-6 Luna. Whether the match is coincidence or strategy, the effect is the same: a pricing floor is hardening at the bottom of the frontier model market, and it's happening fast.
What Haiku 5.5 Actually Is
Anthropica positions Haiku 5.5 as its fastest, cheapest, and most capable small model — the three adjectives that always accompany a new efficiency tier, but rarely all hold simultaneously. The model is designed explicitly for high-volume, cost-sensitive workloads: summarization pipelines, classification jobs, anything where you're running millions of calls and the per-token cost compounds hard.
Output is priced at $0.50 per million tokens — five times the input rate, a ratio consistent with where the rest of the small-model market has settled. The asymmetry reflects reality: output generation is compute-heavier, and providers aren't absorbing that cost.
The more structurally interesting detail: Haiku 5.5 is the first Haiku release to include effort controls. That's a meaningful capability jump for a tier that was previously stateless on that dimension. Effort controls let developers dial compute at inference time — spending more tokens on hard reasoning tasks, less on trivial ones. Bringing that to a sub-$0.10 input-cost model changes the calculus for builders who've been routing between a cheap fast model and a slower expensive one depending on task complexity. Haiku 5.5 starts to collapse that routing decision into a single endpoint.
The $0.10 Anchor and What It Means
Pricing at $0.10 per million input tokens isn't just a number — it's a signal about where Anthropic thinks the mass-market inference tier lives. The fact that this figure matches the GPT-6 Luna input price makes it a two-player bracket: whatever floor OpenAI sets for its smallest frontier-adjacent offering, Anthropic is meeting it exactly.
For operators building on either platform, the implication is direct. When two labs with frontier ambitions converge on the same input price for their efficiency tier, that price becomes load-bearing infrastructure for how developers budget AI into products. Margin math, feature gating, rate-limit strategy — all of it gets recalculated around a number that's now defended from two directions.
The more interesting question is what Anthropic gives up at $0.10. Small-model releases are always a bet that volume covers margin. Haiku 5.5 is aimed at the workloads where that bet is most plausible — classification and summarization are high-frequency, relatively short-context tasks. Long-context, multi-step reasoning is where cost assumptions break down. Haiku 5.5 with effort controls nudges toward the latter category, but how far that goes depends on what developers actually build with it.
The Efficiency Tier Is Becoming the Default Tier
There's a pattern worth naming. The efficiency tier — small, fast, cheap — used to be a fallback. Builders reached for it when the flagship was too expensive or too slow. That positioning is eroding.
When a small model ships with effort controls, the capability gap between it and the flagship narrows for a wide band of real tasks. Haiku 5.5 isn't being sold as a compromise — Anthropic is explicit that it's the most capable model in the Haiku line. For the use cases it's designed for, the argument is that it's not a fallback at all; it's the right choice.
The broader shift here is one of architectural confidence. Labs are investing real engineering in their small models, not just quantizing their flagships and calling it a tier. Effort controls at the Haiku level are a signal that Anthropic is treating $0.10-per-million-input compute as a serious product surface, not a loss leader.
For founders and operators: the floor of what's possible at low inference cost just moved up again. Budget accordingly.