one year on
Anthropic releases Claude Haiku 4.5, a small model it says matches Sonnet 4 on coding at one-third the cost
The five-month-old frontier keeps getting compressed into cheaper silicon, and Anthropic wants Haiku to be the sub-agent workhorse beneath Sonnet 4.5.
Originally published Published on this site
Anthropic on Wednesday released Claude Haiku 4.5, the newest version of its smallest model, available immediately to all users and priced at $1 per million input tokens and $5 per million output tokens. The company says it delivers coding performance similar to Claude Sonnet 4 — the frontier model of five months ago — at one-third the cost and more than twice the speed, and that it surpasses Sonnet 4 at some tasks, including computer use.
Internally, Haiku 4.5 scores 73.3% on SWE-bench Verified and about 41% on Terminal-Bench, roughly level with Sonnet 4, GPT-5 and Gemini 2.5 on those measures and below Sonnet 4.5, which remains Anthropic’s flagship two weeks after its own launch. Anthropic rates the model ASL-2, one safety tier below Sonnet 4.5 and Opus 4.1, and says it showed the lowest rate of misaligned behaviors in the company’s automated alignment assessments to date.
The pitch to developers is orchestration: Sonnet 4.5 breaks a problem into multi-step plans, then dispatches teams of Haiku 4.5s to execute subtasks in parallel. It is available through the Claude API, Claude Code, Amazon Bedrock and Google Cloud’s Vertex AI, where Anthropic bills it as a drop-in replacement for Haiku 3.5 and Sonnet 4.
The record
co-founder of Augment, said Haiku 4.5 hits a sweet spot his firm had not thought possible, reaching 90% of Sonnet 4.5's agentic coding performance at far lower cost
Anthropic's chief product officer, told CNBC the model punches above its weight and that he has begun defaulting to it on the mobile app because answers arrive faster, even if Sonnet reasons better
CEO of Zencoder, said the model unlocks an entirely new set of use cases, running up to four to five times faster than Sonnet 4.5 at a fraction of the cost
distinguished product manager at GitHub, said early testing shows efficient code generation for Copilot with quality comparable to Sonnet 4 at faster speed
One year later — open only if you can handle spoilers
Haiku 4.5 proved unusually durable: in September 2026, nearly a year on, it was still the current Haiku in Anthropic's lineup, listed beneath Sonnet 5 and Opus 5.5 while the larger tiers turned over several times.
Filed
Compiled and edited by Adam Allcock.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy