Step-5-Preview Tested: A Rocket Game With No Rocket

StepFun's Step 5 Preview landed on OpenRouter on 8 October, and I ran it through my VisualForge benchmark the same day. Twice, actually: once through OpenRouter, and again through my OpenCode Go subscription for the two tests that failed. Every page is scored in a browser, and every result is public, so you can open the pages and judge them yourself. Through OpenRouter vs through OpenCode Go: https://visualforgebench.com/compare/?a=20261008T144524_stepfun-step-5-preview_e97b&b=20261008T184320_step-5-preview-free_e330 Step 5 Preview vs Claude Sonnet 5.5: https://visualforgebench.com/compare/?a=20261008T144524_stepfun-step-5-preview_e97b&b=20261005T180131_claude-sonnet-5-5_2fcd Leaderboard: https://visualforgebench.com/board/ All results: https://visualforgebench.com What Artificial Analysis says (checked 9 October 2026): 600 billion parameters, 27 billion active, a 1-million-token context window, text and image input, reasoning, released 18 September 2026. Intelligence Index 44, 40th of 226 models (median 26). $1.00 per million input tokens and $2.70 per million output, about $1.03 per Index task, at about 87 tokens per second. The catch: it wrote 160 million output tokens to run the Index, against a median of 81 million. Correction: in my recording I called it open source. Artificial Analysis lists it as proprietary, so you use it through an API. My results through OpenRouter: 6.5 out of 10, the average of six of the eight core pages, because Alex Optics and Afterburner failed. That's 6th on the board. Two of its scores, Rotor Rampage 7.7 (its best) and Hangar Planner 7.2 (14 of 16 bugs fixed, all 14 features kept), match Claude Sonnet 5.5 on high effort, although Sonnet's overall score is higher at 7.3. Its worst is Launch Day, 4.7: a rocket game with no rocket. Its own wildcard, Cursive Ink Grand Prix, scored 6.4 and isn't in the average. What it used: 5 h 16 min, at least 1.07 million output tokens, 77.4 tokens per second. The board estimates at least $6.24 at OpenRouter's prices; my actual OpenRouter bill that day was $6.50, failed attempts included. The OpenCode Go re-run, same model at high effort: Alex Optics 7.4 and Afterburner 7.4, in 2 h 08 min and 252 thousand tokens, at about 33 tokens per second. The subscription covers it, so there's no per-token bill. It's only two pages, so its 7.4 on the board isn't comparable with full runs. A rough average of all eight core pages across both routes is about 6.7. That mixes two routes, so it's my back-of-the-envelope figure, not an official score. Worth knowing before you compare: - The judge is Claude Opus 5.5 on every test. - I think the judge missed the drawing bug that makes Blast Island unplayable, and 4.7 is generous for a launch with no rocket. - The OpenRouter run's time, tokens and cost are lower bounds, because the two failed tests recorded no tokens. 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 0:36 What is Step 5 Preview 3:19 The benchmark 4:12 Time, tokens and cost 5:21 Breakdown and bug fixing 5:57 Aero Lab 6:23 Jet Odyssey 7:00 Blast Island 7:21 Launch Day 7:48 Rotor Rampage 8:30 Wildcard: Cursive Ink Grand Prix 8:58 The 6.5, and my doubts 9:30 OpenCode Go re-run: Alex Optics 10:04 Afterburner 10:47 Verdict 11:23 Close Which model should I test next? Tell me in the comments, and I'll put it on the leaderboard.