Anthropic shipped Claude Haiku 5.5 on October 7, pricing it at $0.10 per million input tokens and $0.50 per million output tokens for any request under 100,000 tokens. That's a 90% cut from Haiku 4.5's $1.00/$5.00, and it lands exactly on the floor OpenAI set for GPT-6 Luna. Above the 100,000-token line, pricing moves to $0.50/$2.50, a 50% reduction from the previous generation. Anthropic estimates average cost to run falls roughly 75% once the updated tokenizer is accounted for, and notes that about 90% of Haiku 4.5 traffic sat in the short-prompt tier.

The pricing symmetry isn't incidental. Ten days after Opus 5.5 and GPT-6 Sol/Luna arrived on the same afternoon, the small-model floor has been collectively rewritten. Frontier labs are no longer competing on the headline model; they're competing on what it costs to run the always-on substrate underneath it, the classification, triage, and subagent work that touches every production request.

The benchmark deltas make the pitch sharper. On GDPval-AA v2.1, Haiku 5.5 posts 1,620 against Haiku 4.5's 735 and GPT-6 Luna's 1,437, trailing only Sonnet 5.5 at 1,840. On the offline subset of OSWorld 2.1 it hits 72.4%, up from 15.7% for its predecessor and well past Luna's 48.9%. Terminal-Bench 4.0 tells the same story: 39.2% at maximum effort, around 20% at the default medium setting, versus 0.0% for Haiku 4.5 and 16.4% for Luna. The effort dial is new to the Haiku line and matters more than the headline score, because it lets a developer decide per call whether to pay for depth.

Customer-reported figures, vendor-sourced and directional, point the same direction. Asana's Aaron Vinh reports "over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn." HubSpot's Ze'ev Klapow reports Haiku 5.5 scored 92.8% averaged over three runs of its CRM evaluation, the best result among the models it tested. Box's Yashodha Bhavnani reports an 11-point lift over Haiku 4.5 at roughly half the latency. The model ships through Anthropic's API, AWS, Google Cloud, and Microsoft Azure under the identifier claude-haiku-5-5, with a June 2026 knowledge cutoff.

The structural read is the one that matters. When the last comparable floor move happened, OpenAI's gpt-3.5-turbo repricing in 2023, the workloads that followed weren't the ones vendors predicted; they were the ones that had quietly been uneconomic the week before. Scoring every inbound lead, summarizing every call, classifying every ticket in real time: these are now line items a small team can actually run. The frontier gets the press. The floor gets the deployments.

Claude Haiku 5.5 Lands at $0.10/$0.50 — 90% Cheaper for Short Prompts

Sources