MiniMax M3 in OpenCode 1.18.35

as run, 1,403 characters. Every result for this prompt

Build a real-time 3D diorama in Three.js of a small, cosy island valley with a clear mountain stream running through it (one HTML file, CDN imports only, no external assets or textures; everything procedural). The stream is the hero: shallow, crystal-clear water you can see the pebbled bed through, with animated flow, caustic light patterns on the stones underneath, sparkling sun glints on the surface, foam and small rapids where it runs over rocks, and soft rounded river boulders of varied size partly submerged with wet darkening at the waterline. Around it: lush stylised grass and ferns that sway, a few broad-leaf trees, wildflowers, an arched stone bridge or culvert the stream flows through, stepping stones, and a small timber cottage on the bank whose windows glow warmly at night. Lighting presets in a small on-screen pill menu: Morning, Afternoon, Sunset, Blue hour, Night (night shows moonlight on the water and lit windows); transitions animate smoothly. Camera: a slow cinematic auto-fly along the stream by default, free orbit on drag, and a top-down view toggle. Ambient life: drifting petals or fireflies at night, small fish darting under the surface. Gentle procedural water ambience via Web Audio after the first click, with a mute toggle. A title card in the corner names the place. Target 60 fps on a mid-range laptop; include a quality toggle (low/high). Deliver index.html.

Headline cost $0.35, as reported by OpenCode (its own bill for the run). The list-price estimate from the token counts is $0.35.

Total computed from tokens

$0.3480

Total reported by the harness

$0.3519

Difference

matches

OpenCode 1.18.35 reported its own total; token counts x list prices shown beside it. The harness reports reasoning tokens separately from output; here output includes them (43,999 visible + 25,484 reasoning), and both bill at the output rate.

Prices: https://fireworks.ai/models/fireworks/minimax-m3, 2026-10-08.

one prompt, agentic harness (self-review allowed)skills=on

• frozen prompt (written): core requirements sha256 e84a7975916b, full harness prompt sha256 e84a7975916b; exact prompt text used by the existing harness rows of this test on the runs branch (copied from threejs_island_stream_diorama__claude-code__opus-5-5-high__attempt1/meta.json, sha256 verified); Beta h
• no TUI: the harness ran in its non-interactive mode, so there are no TUI screenshots; the timelapse is built from version renders only
• OpenCode config (providers only, permission allow-all except question; no agents/instructions) + both default skill kits in xdg/config/opencode/skill (skills=on)
• infra retries: 0
• in the published transcript the sandbox HOME directory is shown as <HOME>
• skills=on: 54 skills installed in the harness skill directory before launch (video-33 remotion-dev/skills + heygen-com/hyperframes@0b5db9da + f; web-threejs-21 [redacted]/ai-AI Model Reviewer skills/ (alpha/mr-homebar)@7521074d98c9).
• Mode: headless run mode (opencode run --format json), inside bubblewrap sandbox
• Isolation audit: none
• Renderer: Three.js r160 (three@0.160.0 module plus the OrbitControls addon, loaded from unpkg through an importmap) with WebGLRenderer and the page's own ShaderMaterial shaders; extra 2D canvases created by the page's script; sound through WebAudio.
• Card video: the run's own video.mp4 (H.264 yuv420p, unchanged). Poster: the run's own render.png.
• Skills on: the run had 54 skills installed (33 video + 21 web/Three.js), as its meta records.
• MiniMax M3 served via Fireworks (OpenCode provider fireworks-ai).
• Same frozen prompt as the other rows on this test; configurations differ and all rows are unscored, so this is not a head-to-head result.
• Live demo: not published. The page throws on load in any browser from its own code: index.html line 1189 subtracts two THREE.Vector3 objects with '-' (streamCurve.getPointAt(0.5) - cottagePos), which gives NaN, and the next line (lookStream.y = 0) throws a TypeError that stops the module before the scene is built, so the page stays on its 'tending the moss...' loading screen. The harness reported the run finished, so the status is kept as recorded (ok). The poster is the run's own render.png, which shows that same loading screen.