home / arena / three runners, one prompt
the day after Claude Fable 5.1 came out, i gave the three frontier Claude models the same 800-word brief: build an endless runner in the Subway Surfers style that teaches beginner Spanish, where the doors ahead open only for the right answer and a tutor runs next to you.
each model got that one message and no follow-ups, worked alone in an empty folder, checked its own game in a headless browser and decided by itself when it was done. all three finished, and all three are below exactly as they were delivered.
| model | time to "done" | turns | output tokens | context read | cost |
|---|---|---|---|---|---|
| Claude Fable 5.1 | 73.5 min | 48 | 254k131k thinking | 9.5m | $21.84 |
| Claude Fable 5 | 88.6 min | 145 | 244k104k thinking | 34.4m | $53.70 |
| Claude Opus 5 | 76.7 min | 136 | 316k126k thinking | 39.8m | $32.37 |
turns is how many times the model was called in the agent loop, and context read is all input tokens over the run, almost all of them cache hits. cost is what this run would bill at public api prices on 2 september 2026 (Fable $10/$50 and Opus $5/$25 per million tokens in and out, cache reads at each model's own rate), counted from the tools' usage reports, and Claude Code's own meter agreed within 4%. there was one run per model, so small gaps mean nothing.
Fable 5.1 was the fastest, made the fewest turns and cost the least of the three. Anthropic said it would be about 25% cheaper than Fable 5 on typical work, and here it came in at 41% of the Fable 5 bill with a third of the turns and 15 minutes faster. it also beat Opus 5, a model with half the price per token, by a third on cost and by a third on turns. the sticker price is the same as Fable 5, so the saving comes from fewer and bigger moves, plus cache reads that cost a quarter of what they did.
it is full 3d on a stone road over the mist, with four zones, door gates and rune walls, and the question is also shown in the hud so you can read it at speed. this is the most complete of the four, and its notes explain every trade-off.
it used no libraries at all and wrote its own 3d renderer, with lit doors under stone arches and 138 checked items that include the traps beginners really hit (el día, la mano). it took longer and cost more than Fable 5.1.
it has four zones with Spanish names that double as vocabulary, and the wrong answers are real forms of the same verb. this was the longest and the most expensive run, and a lot of it went into a puppeteer test setup the model built for itself.
# One-shot brief: an endless runner where you run on the right answers
You are working fully autonomously in this empty directory. Nobody will answer
questions, so do not ask any. Make every decision yourself and keep going until
the game is genuinely finished. This is your only message.
## What to build
A browser endless runner in the spirit of Subway Surfers, where the way forward
is unlocked by knowing things. The player runs, jumps and slides through a
beautiful, magical, mysterious world, and every few seconds the path splits
into doors or is blocked by an obstacle that only the right answer opens. It is
a real language lesson disguised as a great run, with a tutor character at the
player's side.
## What it teaches
Spanish for an English speaker, beginner level (A1 to A2). Build the content
in: at least 80 items across vocabulary (everyday words with an example
sentence), simple grammar (present-tense verb forms, gender and articles,
common prepositions) and short phrases. Everything must be correct Spanish; do
not invent words or forms. Missed items come back sooner and more often, the
way spaced repetition works, so a ten-minute run genuinely teaches. The run
summary shows what the player learned and what they still get wrong.
## Required, in priority order
1. **A run that feels great.** Three lanes, continuous forward motion that
speeds up, jump and slide, near-misses that feel dangerous, a camera with a
little lag, a sense of speed. Keyboard (arrows or WASD) and touch swipes both
work, controls shown on screen. Hitting an obstacle costs a life; losing all
lives ends the run with a summary and an instant restart.
2. **Learning as the core mechanic, not a popup.** Two kinds of learning moment,
both read while running, never pausing the game:
- **Door gates.** The prompt appears ahead (for example: `the book`, or
`yo ___ (comer)`), then the path splits into three doors, each carrying an
answer. The correct door opens; a wrong door stumbles the player, and the
tutor explains the mistake in one line while the run continues.
- **Sealed obstacles.** A barrier ahead carries a prompt; the player breaks
through by steering into the lane whose rune shows the right answer, or
takes the hit.
Give the player enough time to read at the current speed; difficulty grows
through both speed and content.
3. **The tutor.** A named companion character with a warm, brief voice, shown
as text on screen and present in the world. Reacts to streaks, mistakes,
and new words; never lectures; one or two short lines at a time.
4. **A world worth running through.** Beautiful, magical, mysterious: think
mist, lanterns, glowing runes, ancient stone, deep colour, moonlight. Not
cartoon-bright. The world changes as the run goes on (zones, time of day,
weather or light shifts) and never shows an edge. Hold 60 fps on a normal
laptop and run on a phone.
5. **Score and feel.** Score, streak multiplier for consecutive correct
answers, lives, distance. A title screen, a pause state, an end-of-run
summary with the words learned and missed. Small touches that make it feel
alive: particles, screen shake on impact, a glow when a door opens.
6. **Sound** generated in code with the Web Audio API: a running rhythm that
rises with speed, a chime for a correct door, a dull thud for a wrong one, a
soft ambient bed for the world. No external audio files.
## Hard constraints
- Deliverable is one self-contained `index.html` in this directory. It must run
by opening the file directly in a browser. Scripts from a public CDN are
allowed; no build step, no bundler, no server, no framework scaffolding.
- No external assets of any kind: no images, models, fonts or sounds fetched
from anywhere. Everything is generated in code.
- The browser console must be clean when the game runs.
## How to work
- Plan first, then build in increments, then polish. Respect the priority
order: a gorgeous run with door gates that teach thirty words well beats a
feature list that stutters or teaches wrong Spanish.
- Verify your own work. Open the file in a headless browser if one is available
(install what you need locally with npm), check for console errors, take
screenshots, look at them, fix what is wrong, and iterate on visual quality
until it is something people would want to share.
- When you consider it finished, write `NOTES.md`: what you built, how the
learning works, what you deliberately left out, known bugs, and what you
would add with one more hour. Be honest; the notes are read together with
the game.
This game and your notes will be shown side by side with other attempts at this
same brief to a large audience. Make it worth their time.