"Anthropic trained Claude Haiku 4.5 on a proprietary dataset combining publicly available internet information through February 2025, non-public third-party data, contributions from data-labeling services and paid contractors, user data from Claude users who opted in, and internally generated data at Anthropic. The company applied multiple data cleaning and filtering techniques during training, including deduplication and classification methods. The model operates as a hybrid reasoning system, allowing users to choose between two response modes."
"By default, Claude Haiku 4.5 answers queries rapidly, but users can activate an "extended thinking mode" where the model allocates additional time to consider its response before answering. This capability represents a departure from Claude Haiku 3.5, the previous model in the small-model class, which lacked any extended thinking functionality. When users receive responses generated through extended thinking mode, they can access the model's reasoning process."
Anthropic released Claude Haiku 4.5, a small, fast hybrid reasoning large language model available to all users. The model aims to match Claude Sonnet 4 performance while costing one-third as much and running over twice as fast, with particular effectiveness on coding tasks and computer use. Training data combined public internet content through February 2025, non-public third-party sources, paid labeling contributions, opt-in user data, and internally generated examples, with deduplication and classification filtering applied. The model offers two response modes, including an extended thinking mode that reveals its reasoning process, which the company cautions may have uncertain faithfulness.
Read at InfoQ
Unable to calculate read time
Collection
[
|
...
]