Anthropic launches Claude Haiku 5.5 small model with 75 percent lower costs

Haiku 5.5 is Anthropic leaning hard into the economics of AGI. Frontier models grab headlines, but most real workloads are short, repetitive and highly cost sensitive. By cutting Haiku’s effective price by 75 to 90 percent while keeping a 1M token context and solid benchmark scores, Anthropic is trying to make it irrational for enterprises to run these jobs on anything else. That is a direct play for the background traffic that will fund longer term frontier research.

Strategically, this is also an agent story. Anthropic is explicitly positioning Haiku 5.5 as the sub‑agent that handles high volume tasks while Opus and Sonnet focus on harder reasoning. The release notes, customer testimonials and SDK updates around browser and computer use show they are optimizing a whole stack for agentic workloads, not just a single model. That puts pressure on OpenAI, Google and Meta to match both performance and unit economics in the “small but smart” tier.

For the broader race to AGI, cheap, fast models like Haiku 5.5 are the substrate that lets teams spin up thousands of agents, run massive tool‑using experiments and collect interaction data at scale. The frontier breakthroughs will still come from the biggest models, but it is this layer that makes sustained experimentation and deployment financially viable.