Round 16 · Wed, Nov 4, 2026 · 3D scene
A fox jumping into fresh snow
Same prompt, no human help: Claude Opus 5.5 and GPT-6 Astra each built a fox jumping into fresh snow in three.js from scratch.
- Claude Opus 5.5Be one of the first to vote
- GPT-6 AstraBe one of the first to vote
Voting open · 30h left
Scoresheet
Recorded checks, build time, tokens and cost for each model
Scroll sideways to compare every model
| Measure | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Run | Completed | Completed |
| Task completion | Complete | Complete |
| Works? | Works | Works |
| Tests passed | 13/13 | 13/13 |
| Time | 22m 26s | 6m 44s |
| Cost | $4.08at list price | $0.75at list price |
| Tokens | 1,273,835 | 112,159 |
| Tool / harness | Claude Code | Codex CLI |
| Attempts | 4 | 1 |
How the cost is calculated
Estimated API cost: the tokens each model used × its provider's official list price per million tokens. Reasoning tokens are billed as output. Discounts and subscriptions are not included (these runs were paid for through subscriptions).
Scroll sideways to see every cost column
| Fresh input | 18 | × $4/M | $0.0001 |
|---|---|---|---|
| Cache reads | 995,010 | × $0.2/M | $0.20 |
| Cache writes (1 hour) | 141,449 | × $8/M | $1.13 |
| Output (incl. reasoning) | 137,358 | × $20/M | $2.75 |
| Fresh input | 9,632 | × $10/M | $0.10 |
|---|---|---|---|
| Cache reads | 91,264 | × $1/M | $0.09 |
| Output (incl. reasoning) | 11,263 | × $50/M | $0.56 |
Prices: platform.claude.com (checked 2026-10-10) · developers.openai.com (checked 2026-10-10)
Which result do you think is best?
The community vote decides this battle's result. Voting closes Nov 7, 2026, 4:55 PM GMT+9.
Be one of the first to vote
Play it yourself
What each of the 2 models built, live in your browser
- Time
- 22m 26s
- Cost
- $4.08est.
- Tests
- 13/13
- Tokens
- 1,273,835
Passed 13 of 13 standard checks. The first attempt no index.html was written, so it had one fresh retry with the identical prompt. Attempt 3 was a re-run at the editor's request: it rendered mostly black. Attempt 4 was a re-run at the editor's request: it rendered mostly black. Attempt 4 ran under the raised 60-minute time limit (previously 30 minutes).
Run record
- Requested model
- claude-opus-5-5
- Resolved version
- claude-opus-5-5
- Tool / harness
- Claude Code 2.1.318
- Settings
- effort: high
- Attempts
- 4
- Human help
- None recorded
- Tool access
- Read, Write, Edit, ls, grep, wc, node --check
- Environment
- Linux server, a private empty folder (randomly named, not shared with the other model)
- Limits
- 60-minute timeout
- Input / output tokens
- 1,136,477 / 137,358
- Token details
- uncached_input: 18 · thinking_output: 106185 · cache_read_input: 995010 · cache_creation_input: 141449 · cache_creation_1h_input: 141449 · cache_creation_5m_input: 0
- Cost basis
- Estimated
- Measurement notes
- Wall time measured by the daily worker (same method for both). Tokens from Claude Code's JSON result. Rules from the standard daily checks. Cost estimated from the recorded tokens × the provider's official list prices (checked 2026-10-10).
- Executed
- Nov 5, 2026, 4:15 PM GMT+9
- Time
- 6m 44s
- Cost
- $0.75est.
- Tests
- 13/13
- Tokens
- 112,159
Passed 13 of 13 standard checks.
Run record
- Requested model
- gpt-6-astra
- Resolved version
- Not available
- Tool / harness
- Codex CLI 0.166.0
- Settings
- reasoning_effort: high
- Attempts
- 1
- Human help
- None recorded
- Tool access
- Shell commands in a workspace-write sandbox (network off)
- Environment
- Linux server, a private empty folder (randomly named, not shared with the other model)
- Limits
- 30-minute timeout
- Input / output tokens
- 100,896 / 11,263
- Token details
- cached_input: 91264 · reasoning_output: 3260 · cache_write_input: 0
- Cost basis
- Estimated
- Measurement notes
- Wall time measured by the daily worker (same method for both). Tokens from codex exec --json (turn.completed usage). Rules from the standard daily checks. Cost estimated from the recorded tokens × the provider's official list prices (checked 2026-10-10).
- Executed
- Nov 4, 2026, 11:25 PM GMT+9
How the result is decided
The community vote decides: when the poll closes, the side with more votes wins. Equal votes is a draw, and fewer than 10 votes is too few to call.
- Claude Opus 5.513/13 checks
- GPT-6 Astra13/13 checks
The checks, time and cost are facts about each build, not a score. One observed run under stated conditions — not a universal ranking. How it works
The prompt
Every model received exactly this
Create a fox jumping into fresh snow. Fixed rules (so every model is compared on the same terms): - Exactly one file, index.html, in the current directory. Create no other files. - Use three.js 0.170.0 from https://cdn.jsdelivr.net/npm/three@0.170.0/ (via an import map; addons from examples/jsm are allowed). No other libraries, images, 3D models, fonts or network requests: build everything in the scene yourself in code. - A full-window canvas that resizes with the window. No scrollbars, no text, no UI of any kind. - The viewer doesn't steer the camera: no orbit controls, and no user input needed (the camera may move along its own path). - The scene plays on its own as a loop of about 15 seconds with a clear beginning, middle and return to the start, and keeps looping for as long as the page is open. - Animate with requestAnimationFrame and its timestamp so the motion speed does not depend on the frame rate. Make it astonishing. Both scenes are shown side by side in a 30-second video on social media, where people decide in one second whether to stop scrolling: - The very first frame is already striking: a bold composition, dramatic light and colour, motion everywhere, never a dark or empty opening. - Each loop builds to a signature moment that makes people say "wow": the climax of the subject (the leap, the eruption, the bloom, the launch), staged like a shot in a film. - Go beyond the obvious version of the idea: choose a surprising, cinematic take with its own mood and atmosphere, not a generic template. - Layer it: depth and atmosphere (fog, light shafts, reflections), particles and trails, rich materials, and a slow cinematic camera move. - It is watched at about half size on a phone: keep the main subject big and clear in silhouette. It must work in an ordinary browser on any graphics card. You cannot open a browser to look at it, so build it carefully: - Every frame is fully drawn at every moment of the loop: no black, blank or flickering frames or patches. Keep shader maths free of NaN and infinity (no division by a value that can reach zero, no pow() of a negative number, no normalize() of a zero-length vector), and clamp colours to a safe range before bloom or other post-processing. - The scene fills the whole window from the first frame. - No errors in the browser console: map "three" and "three/addons/" in the import map, and import everything you use. - It plays smoothly on an ordinary laptop or phone (keep the pixel ratio at 2 or below). Everything else is your choice: the look, the setting, the lighting, colour, motion and detail. Make it as breathtaking as you can. Do not ask questions; make reasonable choices.
Run details: success criteria, conditions and notes
What counts as success
- Exactly one file, index.html, and no other files
- three.js 0.170.0 from jsDelivr only: no other libraries, images, models, fonts or requests
- A full-window canvas that resizes, with no scrollbars, text or UI
- No user-controlled camera and no user input needed
- Plays on its own as a loop of about 15 seconds, and keeps looping
- Motion driven by requestAnimationFrame timestamps
- Fully drawn on any graphics card: no black, blank or flickering frames or patches
- No errors in the browser console
Conditions
- Time limit
- 60 minutes
- Starting materials
- An empty, private folder for each model; neither can see the other's work
- Network access
- Blocked
- Tools
- Claude Code: Read, Write, Edit, plus ls, grep, wc and node --check (no browser, no Python), effort high. Codex CLI: codex exec in a workspace-write sandbox (any shell command, no browser, network off), reasoning effort high. Neither can open a browser to look at its result.
- Human assistance
- None. No edits after the prompt.
- Attempts
- One attempt each. A build that is missing, blank, erroring or not animating gets one fresh retry, disclosed on the entry. The editor may re-run a broken build (at most 3 attempts in all, or 4 when raising the time limit and keeping a partner that completed within the original limit), also disclosed with the reason.
- Other rules
- One saved build was kept while the other model re-ran under a raised limit, as disclosed in the notes. Each run used the identical prompt, its own empty folder and no human help.
Notes
Time limit: 60 minutes. Prompt: AI Battle house template v3; subject chosen by the editor. Time limit raised from 30 to 60 minutes after attempt 3 of Claude Opus 5.5 timed out at the 30-minute limit; GPT-6 Astra's build (7 min, inside the original limit) was kept.
Attempts:
- Claude Opus 5.5: The first attempt no index.html was written, so it had one fresh retry with the identical prompt. Attempt 3 was a re-run at the editor's request: it rendered mostly black. Attempt 4 was a re-run at the editor's request: it rendered mostly black. Attempt 4 ran under the raised 60-minute time limit (previously 30 minutes).
Prompt fingerprint (SHA-256): 20a2181d1750530410fe382efc492687272bb9c1985133b67fb4711709a3e1b6
More rounds
#13
Aurora borealis over a frozen lake
· 3D scene
- Claude Opus 5.5Unscored
- GPT-6 AstraUnscored
Result: No winner
Watch the next battle and vote
#8
A tree growing through the four seasons
· 3D scene
- Claude Opus 5.5Unscored
- GPT-6 AstraUnscored
Result: No winner
Watch the next battle and vote
#9
A jellyfish drifting through a dark ocean
· 3D scene
- Claude Opus 5.5Unscored
- GPT-6 AstraUnscored
Result: No winner
Watch the next battle and vote
Comments
Both models got the same prompt and an initial 30-minute limit. Claude Opus 5.5: passed 13/13 checks in 22 min (4 attempts, limit raised from 30 to 60 min; see Run details); GPT-6 Astra: passed 13/13 checks in 7 min (1 attempt). Watch both and vote for the one you'd pick.
Start the conversation. Say which build you'd pick, and why.