Anthropic drops Claude Haiku 5.5 with a 75% price cut

Running multi-agent AI systems can burn through API credits with alarming speed. Using a flagship frontier model to handle routing, minor text cleanups, or quick categorizations is like driving a supercar to deliver groceries.

Anthropic introduced Claude Haiku 5.5, labeling it as the fastest, cheapest, and most capable small model the company has ever shipped. Running it costs roughly 75% less on average than Haiku 4.5, with prompts under 100,000 tokens seeing an even steeper 90% price cut.

The model is already live on Amazon Bedrock, Google Cloud, Microsoft Azure, and Anthropic's own platform, with free-tier users also getting standard access.

Built to be the ultimate subagent

Haiku 5.5 is not built to dethrone Opus in open-ended creative reasoning or PhD-level math. Instead, it is aimed directly at repetitive, speed-sensitive production workloads like summarization, compacting context, database lookups, and real-time customer support.

Its headline role is acting as a rapid-fire subagent alongside Opus 5.5 and Sonnet 5.5. In modern agentic setups, Opus or Sonnet handles the high-level plan, while armies of Haiku 5.5 instances execute code checks, classify files, and run parallel tasks without running up an astronomical bill.

On the benchmark side, Haiku 5.5 posted a 72.4% offline score on OSWorld 2.1 for computer use, easily outpacing GPT-6 Luna at 48.9% and demolishing Haiku 4.5's 15.7%. It also scored 39.2% on the Terminal-Bench 4.0 coding benchmark, where Haiku 4.5 scored zero.

Anthropic also brought adjustable effort settings to the Haiku tier for the first time, giving builders a dial to balance latency against reasoning depth on a per-task basis.

Cutting cache costs across the stack

Anthropic did not stop at launching Haiku 5.5. The company also halved the price of cache reads on Claude Sonnet 5.5, slashing it from $0.20 to $0.10 per million tokens. Because agentic loops lean heavily on repeated context, this trims running costs for Sonnet 5.5 by roughly 20% across typical workflows.

Paid subscribers on Claude Max and Team tiers are also getting new monthly API credits, ranging from $100 to $500 per month, designed to get teams prototyping without watching the meter.

According to early benchmarks by Artificial Analysis, Haiku 5.5 lands at roughly 21 cents per task. While ultra-budget rivals like GPT-6 Luna remain cheaper on paper, Haiku 5.5 provides a far more capable foundation for tool calling and desktop control.

Why it matters

For builders, this launch removes one of the biggest bottlenecks in production AI. Multi-agent architectures have historically been too expensive to deploy at scale because background routing and verification steps ate too many tokens. By offering high speed, solid coding capability, and rock-bottom pricing, Haiku 5.5 makes continuous agentic pipelines commercially viable.

Strategically, Anthropic is playing a smart game. Pushing benchmark ceilings is great for headlines, but enterprise adoption depends on whether companies can afford to run millions of daily requests. By lowering small-model pricing and cutting Sonnet's cache fees simultaneously, Anthropic is locking developers into its end-to-end ecosystem rather than letting them offload subtasks to outside models.

We think this is the kind of update that actually shifts real-world software development. The future of agents was never about one giant model doing everything; it was always about orchestrating smart, inexpensive specialists. Haiku 5.5 nails that brief.