AI Model Reviewer

DeepSeek V4.1 Flash in dsh 0.1.5-rc.3

skills off

← Open in the gallerySame prompt, all models

Zoom, inline player and more ways to view
Poster frame of DeepSeek V4.1 Flash's page
Open the original page

The model's own page, run in a sandboxed frame only when you ask (a page that needs its own origin opens in a new tab on a separate demo site). Phones get a full-screen player capped at 1.5x pixel density (2x on larger screens) so it stays smooth; the original page runs uncapped. It starts in its auto-tour (T switches to play mode on a keyboard).

OKJellyfishmax effortNeutral harness

Wall time 45m 32sCost $0.26 estimated

Snapshot 4 of 4 saved during DeepSeek V4.1 Flash's run
  1. Snapshot 1 of 4 saved during DeepSeek V4.1 Flash's run
  2. Snapshot 2 of 4 saved during DeepSeek V4.1 Flash's run
  3. Snapshot 3 of 4 saved during DeepSeek V4.1 Flash's run
  4. Snapshot 4 of 4 saved during DeepSeek V4.1 Flash's run
Still snapshots saved during the run (no motion recording): 4 distinct frames.

Generation timelapse

Poster for generation timelapse

Generation timelapse (the agent's screen during the run, sped up). This is not the finished video. Download video (1.4 MB).

Review

No verdict yet.

The full review

OK. Finished on its own after 45.5 min.

Scores

AI judge

Not scored

No rubric for this prompt

Visitor votes and reviewer grades are not available yet.

AI judge
AI judge: the test runner (model claude-opus-5-5), scoring by visual review of the 1600 px render against the pre-registered rubric. No AI-judge score exists for this result: there is no rubric for this prompt, so no score is shown and none is made up.
Visitors
Public voting with a blind reveal is coming: you score the output first, then see the other scores.
Reviewer
The reviewer's own grade appears here once it is added.

Scores compare results for the same prompt only. AI-judge boards rank runs within one prompt, equal totals sharing a rank; scores from different prompts are never averaged.

Downloads

Files actually published for this run. Original output, site-made previews and post-run conversions are labelled separately.

The exact prompt

as run, 1,367 characters. Every result for this prompt

Build a Three.js scene of a huge glowing jellyfish walking through a rainy city at night on very tall, thin, jointed legs, presented as a live disaster news broadcast. The jellyfish's translucent bell pulses with soft bioluminescent light, its tentacles sway beneath it, and each slow stride steps between the towers while rain streaks through the light and wet streets reflect neon signs and headlights. Frame it like a TV emergency broadcast: a main camera feed plus several smaller feeds from other angles (street level, a helicopter, a rooftop) that update as the creature moves; a HUD with a channel logo, a red LIVE badge, a scrolling news ticker, a clock and coordinates; and an event log panel that adds timestamped entries as the jellyfish crosses the city. Keep the mood tense and cinematic, with fog, flickering streetlights and flashing red and blue emergency lights. Everything is built in code from Three.js geometry, shaders and materials; no external models or textures. Build it as index.html (plus any local asset files you create) in the current directory. Three.js may load from a pinned CDN. You may serve it locally and open it in headless chromium (/usr/bin/chromium, software WebGL, no GPU) to screenshot and inspect your own output, and iterate on your own before finishing. Close any browser you open as soon as you've taken your screenshot.

How it was made

Wall time 45m 32sFirst artifact at 1m 53s9 saved versions

Versions saved during the run

9 intermediate outputs, in the order the model wrote them.

Step by step

The prompt, every agent turn and tool call (tool name and a one-line summary; inputs and outputs are not shown), then the final message. The full transcript is 4.4 MB; it is published as a scrubbed, compressed download rather than shown on the page.

Show all 330 steps

Loading the timeline…

Download the full transcript dsh session log, 925 KB, gzip. Scrubbed: anything that looks like a key, token, e-mail address or phone number is replaced, and account details and the test machine's file paths are removed.

Cost breakdown

Headline cost $0.26, estimated from the token counts at list price.

Tokens by typeTokens$ per 1MCost
Input (uncached)91,653$0.15$0.0137
Cache write0$0.15$0
Cache read32,658,560$0.00$0.0980
Output239,760$0.60$0.1439
of which reasoning111,990billed as output
Total computed from tokens
$0.2556
Total reported by the harness
none

Computed from harness-reported token counts x list prices. DeepSeek bills no cache writes (a cache miss is input), so cache_write is 0; dsh prints no dollar figure.

Prices: https://api-docs.deepseek.com/quick_start/pricing, 2026-09-24.

Run notes

one prompt, agentic harness (self-review allowed)prompt rewritten after the 2026-09-24 box losslabelled re-run (of threejs_jellyfish__dsh__deepseek-flash)headless mode (TUI not automatable)

  • frozen prompt (rewritten): engine/prompts_rewritten/threejs_jellyfish.txt: written by us from the recorded description (zero results)
  • dsh has no terminal UI: headless mode (TUI not automatable), run in tmux for screenshots
  • dsh session log decompressed from session.v3.jsonl.zstd
  • Mode: headless mode (TUI not automatable) — dsh --profile headless in tmux, inside bubblewrap
  • Isolation audit: outside paths seen