And talking of price, Haiku 4.5 is included for subscribers of the Claude net and app plans. By means of the API (for builders), the small mannequin is priced at $1 per million enter tokens and $5 per million output tokens. That compares to Sonnet 4.5 at $3 per million enter and $15 per million output tokens, and Opus 4.1 at $15 per million enter and $75 per million output tokens.
The mannequin serves as a less expensive drop-in substitute for 2 older fashions, Haiku 3.5 and Sonnet 4. “Customers who depend on AI for real-time, low-latency duties like chat assistants, customer support brokers, or pair programming will recognize Haiku 4.5’s mixture of excessive intelligence and noteworthy pace,” Anthropic writes.
Claude 4.5 Haiku solutions the traditional Ars Technica AI query, “Would the colour be referred to as ‘magenta’ if the city of Magenta didn’t exist?”
On SWE-bench Verified, a check that measures efficiency on coding duties, Haiku 4.5 scored 73.3 p.c in comparison with Sonnet 4’s comparable efficiency stage (72.7 p.c). The mannequin additionally reportedly surpasses Sonnet 4 at sure duties like utilizing computer systems, in keeping with Anthropic’s benchmarks. Claude Sonnet 4.5, launched in late September, stays Anthropic’s frontier mannequin and what the corporate calls “the most effective coding mannequin obtainable.”
Haiku 4.5 additionally surprisingly edges up near what OpenAI’s GPT-5 can obtain on this specific set of benchmarks (as seen within the chart above), though because the outcomes are self-reported and probably cherry-picked to match a mannequin’s strengths, one ought to at all times take them with a grain of salt.
Nonetheless, making a small, succesful coding mannequin might have surprising benefits for agentic coding setups like Claude Code. Anthropic designed Haiku 4.5 to work alongside Sonnet 4.5 in multi-model workflows. In such a configuration, Anthropic says, Sonnet 4.5 may break down complicated issues into multi-step plans, then coordinate a number of Haiku 4.5 situations to finish subtasks in parallel, like spinning off staff to get issues performed quicker.
For extra particulars on the brand new mannequin, Anthropic launched a system card and documentation for builders.
