Haiku Starts to Lag Behind in Hallucinations and Cost-to-Performance

Claude Haiku used to be the preferred choice for fast, cheap tasks.

A Jump in Fabrications in Typical Deployments

According to community feedback, that is changing. Developers are complaining about a significant increase in hallucinations in typical deployments.

The Budget Category Shifts the Baseline

The criticism isn't happening in a vacuum. While Haiku stagnates, new iterations of budget models have hit the market, most notably GPT-5.6-Luna, offering more accurate outputs for the same price. Haiku has lost its main competitive advantage.

A Hidden Risk for Claude Code Tools

The issue is compounded by reports that Haiku may still be the default model for internal tools within the Anthropic ecosystem, such as WebFetch in Claude Code. If a foundational web reading tool hallucinates, it threatens the entire developer workflow.

The Next Generation Will Decide It

The current situation shows that the small model category forgives no hesitation. Anthropic will need to dramatically raise the bar with their next release, or risk losing the developer community in this segment.

Lilith's verdict

When a small model fabricates facts inside a web reading tool, it's like having an assistant rewrite the newspaper before putting it on your desk.