Claude Haiku 5.5 vs GPT-6 Luna

Claude Haiku 5.5 and GPT-6 Luna charge the same price per token, so I gave them the same eight builds in my VisualForge benchmark: a wind tunnel, a jetpack game, a Worms-style battle game, a Starship-style rocket flight, a camera shop website, a jet engine, a bug-fixing app and a helicopter shooter. Haiku also designed its own wildcard test. Every page is scored in a browser, and every result is public, so you can open the pages and judge them yourself. Compare the two runs: https://visualforgebench.com/compare/?a=20261002T212950_gpt-6-luna_9fb0&b=20261007T191801_claude-haiku-5-5_b208 All results: https://visualforgebench.com Average score over the eight core tests: GPT-6 Luna 6.2, Claude Haiku 5.5 6.1. A draw, in my book. Haiku's wildcard, Cold Shock Parison, scored 7.15 and isn't in the average. The bill is where they split. Haiku wrote at about 222 tokens per second, five times faster than Luna's 44, and wrote at least 2.58 million output tokens against Luna's 577 thousand. At API prices that's at least $3.79 for Haiku and $0.85 for Luna. Time was close: at least 4 h 14 min for Haiku, 3 h 38 min for Luna. Prices: both list at $0.10 per million input tokens and $0.50 per million output (OpenRouter, 8 October 2026). Anthropic charges $0.50 and $2.50 for Haiku prompts over 100k tokens. Worth knowing before you compare: - The judge is Claude Opus 5.5, so Anthropic's big model is marking its little sibling's homework. - Luna ran at max effort, Haiku at high. A Haiku max-effort run is next, and its results will go on the website. - Haiku's time, tokens and cost are lower bounds: six of its attempts hit the 128k output-token limit and recorded no tokens. - Some Haiku attempts failed or were cut short, so I re-ran those tests and the best attempt counts. Luna had one attempt at each test. - Luna ran on the older briefs, which don't ask for sound. - The bug-fixing test used different apps: Luna fixed Hangar Ops, Haiku the bigger Hangar Planner. - Repaired pages are shown for viewing only. The scores come from the original code. - Both runs used my subscriptions. Costs are what the same work would cost at API prices. I recorded my commentary while going through the results. The narration is a tightened script of what I said, read by my own HeyGen voice clone. The background music was made with Google Gemini. 0:00 Intro 0:42 What Anthropic released 1:58 The benchmark 3:03 Time, tokens and cost 4:13 Breakdown and compare 5:23 Aero Lab 6:25 Jet Odyssey 7:36 Blast Island 8:12 Launch Day 10:15 Alex Optics 11:01 Afterburner 11:39 Bug fixing 12:23 Rotor Rampage 13:13 Wildcard: Cold Shock Parison 13:40 Verdict 14:13 Close Which model should I test next? Tell me in the comments.