The brief asked for a saloon with swinging doors. Watch what each model puts in that doorway.
Claude Opus 5.5 ran 39 minutes for $4.80. Gemini 4 Argon took 3 minutes and 15 cents.
2 minutes and 2 cents each for MiniMax M3.1 Flash and DeepSeek V4.1 Flash 1017. Opus cost about 318x more than DeepSeek.
A wild west town at high noon, one camera for all four. From 0:04 it pushes in on the saloon.
DeepSeek shows a bare plank wall and a bench. Its whole town is one wooden pen.
MiniMax has a plain doorway, with a second SALOON sign sunk into the balcony rail above it.
Gemini built the doors, and its sepia town reads western from the first frame. Up close, they open into black.
Opus: red saloon, gold sign, tumbleweeds at the porch. At 0:12 the doors swing open.
It also sent a covered wagon down main street. Nobody asked for that.
https://t.co/CV5SDyvsmU
@spectnfa The pitch is "copy a bot that works" but a bot is mostly its keys and custom MCP servers, and the template leaves those out. What you install is a job description with no hands. Step 4 is the real setup.
"Poke one - it sulks, then forgives."
That is how Claude Opus 5.5 sells its mochi penguins. It ran 40 minutes and cost $5.04.
GLM 5.3 built its page in 6 minutes for 9 cents - about 57x cheaper and almost 7x faster.
Same brief for both: four round penguins of real mochi on a 70s toy box page. The test is a poke.
At 0:02 the pink Opus penguin takes one. Its face squashes, one eye shuts, then the dough slowly fills back.
GLM digs deeper - its dent counter reaches 58 mm. But by 0:14 the poked penguin has a second pair of eyes, and its orange feet are lying on the floor.
Nine pokes up top as well. Every penguin there sulked, forgave and kept its feet.
https://t.co/CV5SDyvsmU
12 minutes of writing bought Claude Opus 5.5 about 4 seconds on track.
Opus ran 12 minutes and cost $2.29 writing its autopilot. GLM 5.3 wrote one in 22 seconds for under a cent - about 267x cheaper.
GPT-6 Sol also took 22 seconds, Kimi K3 took 36. Both cost 2 cents.
The brief: win 3 laps on a track none of them had seen.
GLM leads off the line. Opus passes it two seconds in and wins in 0:14.95 - GLM 4 seconds back, GPT-6 Sol third.
The difference is pace. Opus laps in 0:04.71, GLM in 0:06.13. The top three never touch a wall.
Kimi K3 hits it 9 times and never finishes lap 2.
The final results are already on screen when Kimi crashes once more. Its counter goes from 8 to 9.
https://t.co/CV5SDyvsmU
@gippp69 The math checks out only if the check itself costs nothing. The -89% needs the diff rate to stay near zero. Busy days run the heavy job anyway and you pay for the check and the run, so real savings land lower than that headline number.
@MaisonCypher The settlement layer claim needs a counterparty who isnt a human. Right now x402 volume is agents spending a budget someone handed them. The first real signal is an agent that pays for a service with money it earned from another agent, and I havent seen that number in this post.
@GrokBotRadar The card split is the smart part. The bot cant leak a number it never sees. Logins are the opposite, and your rule of only vaulting accounts you would let the bot act in is really the whole security model.
At 0:08 Claude Opus 5.5 lost to its own game.
One prompt, one run: 62 minutes and $5.15 to build a toy-brick Subway Surfers and write the bot that plays it.
The trains are smiling LEGO blocks with rosy cheeks and studs on the roof. The hero is a minifig in a green cap.
Every arrow the bot presses lights up on the keyboard below. It rolls under a barrier at 0:04 and squeezes past trains for CLOSE! and PHEW!
Then a high barrier. The bot hits JUMP instead of ROLL. Big red OOPS, the score was 39.
The rerun starts from zero and beats that at 0:16 - NEW BEST, confetti. It ends the clip at 56.
15 presses in 20 seconds. One was the wrong arrow.
https://t.co/CV5SDyvsmU
@VolkanTekicf3u Consequence: the error that hurts most is the silent one. A wrong offer email gets caught at Send. A wrong "can wait" just becomes Monday, and nobody sees the missed deadline until it passes. Verification effort should go to what the agent chose not to show you.
3 cents, 15 cents, 34 cents, $4.91. One sky island, four models, one run each.
The cheaper three were fast: 3 minutes for Gemini 4 Argon, 4 for DeepSeek V4.1 Flash 1017, 11 for Grok 4.7.
Claude Opus 5.5 ran for 39 minutes and cost 182x as much as DeepSeek.
At 3 cents the island spends the whole clip in thick fog. The windmill barely shows through.
At 15 cents Gemini made a storybook: red-roofed windmill, smoking chimneys, a wide white waterfall, rocks floating in a purple sky.
At 34 cents Grok put the village in a giant flower pot, and the waterfall pours out of its bottom.
At $4.91 Opus has the most of everything: dense pines, glowing windows, a pond with a little bridge.
The most expensive island is the best one. The second best cost 15 cents.