Claude Haiku 5.5 tested against GPT-6 Luna

Robbert van Empel published a hands-on comparison of Claude Haiku 5.5 and GPT-6 Luna on October 8, 2026. The seven-minute video asks whether Anthropic’s smaller model offers good value for coding and creative projects. It combines intelligence scores and cost-per-task figures with two practical tests: redesigning Wikipedia’s toilet paper orientation page and building a LEGO globe with a South Park theme.

Anthropic introduced Haiku 5.5 on October 7. In its launch announcement, the company positions it for quick, narrowly scoped workloads and as a coding subagent alongside Opus 5.5 and Sonnet 5.5. Anthropic reports an average running-cost reduction of around 75 percent compared with Haiku 4.5. That comparison concerns its previous model; it does not establish that Haiku is cheaper than every competitor.

The Artificial Analysis figures shown in the video give Haiku 43 intelligence points, five more than GPT-6 Luna. However, the cost-per-task comparison puts Haiku at about $0.21, roughly three times Luna’s cost. These figures reflect the evaluation discussed in the video, rather than fixed prices for every project. Token consumption, model settings and the work needed to reach a usable result can change the economics.

In the website test, Haiku produces a dark design with gradients and interactive toilet paper effects. Van Empel finds the styling an improvement over Wikipedia, but considers the animation limited and the remaining page fairly standard. Haiku places eleventh on the leaderboard shown, while Luna is eighth.

The game is more convincing. Haiku creates a recognizable globe, character dialogue and a storyline, although inverted controls and inaccurate brick stacking affect play. It ranks sixth, immediately behind Luna. Van Empel says he would personally reverse those two positions for this particular game, while still favoring Luna overall on price and results.

For makers, the comparison shows why benchmark scores alone cannot settle a model choice. A higher score can coexist with weaker output on a specific brief. Haiku remains an interesting option for people already working with Claude, especially for smaller delegated tasks. The embedded video shows the examples, and The Future’s benchmark lets readers explore the shared briefs and project results.