
"Anthropic launched Claude Haiku 4.5 today. It is the most compact variant of this generation of LLMs from Anthropic and promises to deliver performance close to that of GPT-5. Claude Sonnet 4.5 remains the better-performing model by a considerable margin, but Haiku's benchmark scores are not too far off from the larger LLM. Claude Haiku 4.5 "gives users a new option for when they want near-frontier performance with much greater cost efficiency.""
"It is striking that GPT-5 performs only slightly better than the latest, smallest Claude model in Anthropic's benchmarks. The question is whether this will also be the case in reality. After all, GPT-5 is capable of scaling up to GPT-5 Thinking for complex issues, which yields different benchmark figures. Dual use An interesting suggestion from Anthropic is to use Haiku and Sonnet together."
"Where Sonnet 4.5 costs $3 per million input tokens and $15 per million output tokens, Haiku 4.5 costs $1 and $5, respectively. "Sweet spot" Anthropic further states that Claude Haiku 4.5 is more than just a cheap alternative for users who cannot afford the more expensive models. "Claude Haiku 4.5 hit a sweet spot we didn't think was possible: near-frontier coding quality with tremendous speed and cost efficiency," said Anthropic co-founder Guv Gur-Ari. All told, Haiku achieves about 90 percent of Sonnet's performance, he said."
Claude Haiku 4.5 is a compact variant that aims to deliver performance close to GPT-5 while being far more cost-efficient. Sonnet 4.5 remains the stronger model, but Haiku's benchmark scores are near Sonnet's, achieving about 90% of its performance. Pricing positions Haiku at $1 per million input tokens and $5 per million output tokens versus Sonnet's $3 and $15. A combined approach pairs Sonnet's ability to decompose complex problems with multiple Haiku instances executing the subtasks. Haiku 4.5 behaves more safely than 3.5, with misaligned behaviors more common in Sonnet 4.5 and Opus 4.1, making Haiku the safest Claude model to date.
Read at Techzine Global
Unable to calculate read time
Collection
[
|
...
]