Claude Haiku 5.5: Release, Two-Tier Pricing and Effort Settings

The small model in the Claude 5.5 family

Anthropic released Claude Haiku 5.5 on October 7, 2026, fifteen days after Claude Opus 5.5 and nine days after Claude Sonnet 5.5. It replaced Haiku 4.5 as Anthropic's lowest-cost model. The API ID is claude-haiku-5-5, and the model was offered on Anthropic's platform and through Amazon Web Services, Google Cloud and Microsoft Azure on the day. Anthropic's model documentation lists a 1 million token context window and 128,000 output tokens, and a retirement date no sooner than October 7, 2027. [1][3][2]

Anthropic pitched it for high-volume, latency-sensitive work such as summarization, classification, routing and context compaction, and as a subagent working under Opus 5.5 or Sonnet 5.5. It also said Haiku 5.5 was the fastest model it had released at standard speed, while noting that Opus models in Fast Mode were quicker still. [1][3]

A 90 percent price cut that Anthropic calls 75 percent

For prompts up to 100,000 tokens, Haiku 5.5 cost $0.10 per million input tokens and $0.50 per million output tokens, against Haiku 4.5's $1 and $5. Above 100,000 tokens the rate rose fivefold, to $0.50 and $2.50. Cache reads were $0.01 per million below the threshold and $0.05 above it. [1]

Anthropic's own headline figure was smaller than the list-price cut: the model costs "around 75% less to run" on average. A footnote explains the gap. The token price is 90 percent lower below 100,000 tokens and 50 percent lower above it, and Haiku 5.5 uses slightly more tokens per task because of an updated tokenizer. [1]

Atlas interpretation: The two numbers answer different questions. The per-token price is what a buyer sees on the rate card; the 75 percent figure is Anthropic's estimate of what a typical task costs once the extra tokens and long prompts are counted. A workload that lives mostly above 100,000 tokens would see a much smaller saving than either headline. [1]

An effort dial and Anthropic-reported results

Haiku 5.5 was the first Haiku-class model with an adjustable effort setting; Anthropic's charts show low, medium, high, xhigh and max levels, and its documentation lists medium as the API default. On Anthropic's comparison table, Haiku 5.5 scored 72.4 percent on the offline subset of OSWorld 2.1 against 15.7 percent for Haiku 4.5 and 83.9 percent for Sonnet 5.5, and 39.2 percent on Terminal-Bench 4.0 against 0.0 percent for Haiku 4.5. [1][3]

The same table compares it with GPT-6 Luna, OpenAI's small model, showing Haiku 5.5 ahead on each benchmark where both have a score, for example 72.4 to 48.9 percent on OSWorld 2.1. Customer quotes on the page include Box reporting about half Haiku 4.5's latency in early testing and HubSpot reporting a 92.8 percent average on its internal CRM suite. [1]

Atlas interpretation: Every score above was run or selected by Anthropic, or reported by a customer Anthropic chose to quote. They show what the vendor measured at its chosen settings, not an independent ranking against Luna. [1]

Safeguards and the changes that shipped alongside

Anthropic said Haiku 5.5's cybersecurity safeguards were stricter than Haiku 4.5's but looser than those on its other recent models, and still blocked penetration testing and similar techniques. Its biology safeguards matched Sonnet 5, Sonnet 5.5 and Opus 5. Organizations could apply to Anthropic's Life Sciences and Cyber Verification Programs for broader access. [1]

The same post halved Sonnet 5.5's cache-read price to $0.10 per million tokens, which Anthropic estimated cuts Sonnet 5.5 costs by around 20 percent on most agentic tasks, and began monthly API credits for subscribers: $100 on Max 5x, $200 on Max 20x and up to $500 pooled on Team plans. [1]