Claude Haiku 5.5

Spec

Attribute	Value
Developer	Anthropic
Released	2026-10-07
Announced	2026-10-07
Context window	unknown
Pricing	Prompts up to 100K: $0.10/M input, $0.50/M output; over 100K: $0.50/M input, $2.50/M output
License	unknown
Availability	Claude Platform as claude-haiku-5-5; Amazon Web Services, Google Cloud, Microsoft Azure
Cache reads cost $0.01/M or $0.05/M, and cache writes $0.125/M or $0.625/M, for the same prompt-length tiers. Context capacity and license terms are not established by the captured announcement; the pricing threshold is not a context limit. (source)

Release Date

Available from 2026-10-07. Anthropic describes this as its first Haiku model with adjustable effort. (source)

Benchmarks

Vendor-reported launch results, not independent measurements:

Benchmark	Haiku 5.5
Terminal-Bench 4.0	39.2%
OSWorld 2.1 (offline subset)	72.4%
Humanity's Last Exam (no tools)	45.9%
Humanity's Last Exam (with tools)	57.4%
The table does not attach an effort setting to these Haiku cells. OSWorld's offline subset must stay with the figure; it is not interchangeable with other OSWorld subsets. (source)

Use Cases

Anthropic positions Haiku for summaries, compaction, classification, quick queries and narrowly scoped coding subagents. Its migration guide says the same text uses approximately 30% more tokens than Haiku 4.5. Adaptive thinking is on by default, and the old explicit thinking-budget configuration returns a 400 error; token budgets and integration behavior need rechecking. (source) (source)

Compared To

• Claude Sonnet 5.5 remains the stronger complex-coding option: 70.6% versus Haiku's 39.2% on Terminal-Bench 4.0 in this announcement. (source)
• GPT-6 Luna is a named launch-table comparator. The announcement reports 16.4% on Terminal-Bench 4.0, but the table alone does not establish matched reasoning settings. (source)

Sources

• Launch announcement (source)
• Migration guide (source)