Claude Haiku 4.5 punches above its weight

What counted as absolute frontier intelligence just five months ago has now been shrunk into a lightweight model that runs twice as fast for a fraction of the cost.

Anthropic just launched Claude Haiku 4.5, rolling out its latest small model to all users across web apps, Claude Code, its API, Amazon Bedrock, and Google Cloud Vertex AI.

Frontier smarts at bargain prices

Five months ago, Claude Sonnet 4 was considered state of the art. Today, Haiku 4.5 matches that same level of coding performance while running more than twice as fast at one-third the cost.

On certain specialized benchmarks—like computer use—Haiku 4.5 actually beats Sonnet 4 outright. Anthropic noted this makes browser automation tools like Claude for Chrome noticeably snappier.

Pricing is set at $1 per million input tokens and $5 per million output tokens. That makes it a direct drop-in replacement for both Haiku 3.5 and the older Sonnet 4.

Anthropic is also leaning heavily into agent orchestration. With Sonnet 4.5 released just two weeks ago as the heavyweight thinker, Anthropic suggests letting Sonnet break down complex tasks and commanding a fleet of Haiku 4.5 agents to execute subtasks in parallel.

Safety tests also yielded an unexpected win: in automated alignment assessments, Haiku 4.5 posted lower misaligned behavior rates than both Sonnet 4.5 and Opus 4.1, making it technically Anthropic's safest model to date. It was cleared under AI Safety Level 2 (ASL-2) protections.

Why it matters

We think Haiku 4.5 shifts the economics of real-world AI agents in a massive way.

For developers building real-time coding assistants, customer support bots, or multi-agent swarms, paying top-tier frontier prices has always been the main bottleneck. Getting Sonnet 4-tier coding smarts at $1/$5 per million tokens makes high-volume workflows practical rather than an expensive experiment.

Anthropic’s release pacing is also clever. By locking in the coding crown with Sonnet 4.5 two weeks ago and immediately backing it up with a hyper-fast workhorse, they are boxing competitors out on both peak capability and production-grade efficiency.

Keep in mind that it remains a small model. While great at executing well-defined subtasks and rapid code generation, you will still need a frontier model like Sonnet 4.5 when a project demands deep architectural reasoning.