Opus 5 has completely destroyed my expectations for what an LLM can do. 26 billion tokens. 4 weeks of feeding it Claude accounts nonstop. 124 seconds of 4K, 2978 frames, one camera, zero cuts. Nothing imported. Nothing sculpted by hand. Zero texture files. Every surface is code it wrote... itself. 17,707,774,735 triangles evaluated every single frame. 52.7 trillion across the film. It rented 3 RTX 5090s on its own through middleware another Claude agent built. 231 GPU hours for the final render. The render cost $229. The thinking cost $20,740. The 4K film was the cheap part by 90x. It also rendered the car inside out for weeks and none of us caught it. A back facing solid fills the same silhouette, so it still looked like a car. Audio is the one thing it never got. Rebuilt five to seven times and I reverted to the original every time. A year ago none of this was possible.show more

You can do this with Grok Imagine for $50 in a few hours

Grok 4.6 isn't there yet i tried 4 different times. cant wait to see what grok 4.7 holds.

@elonmusk uh… he’s talking about the video model, not the llm… you spent entirely too much money for that result. 🤯 that had some really awkward camera movement… and was just kinda bland overall.

@elonmusk I mentioned both theres another reply

@joeblowsworth @elonmusk You fucked up and spent $20,000 for no reason lol

@aipulseda1ly @elonmusk i think he’s lying

@tellmechase @elonmusk API equivalent 26 billion tokens. Around 6 and half Claude 20 accounts

So to clarify you spent $20k+ on a 2 minute video?

around 6 full Claude accounts.

So $22-$25k for a 2 minute video. Not including the hardware subscription cost for gpu which was at least $1 a minute? Not video time but compute running. Does it not know how to do this over and over? Can you make more in a few seconds different courses or cars? Or is it just this video?

the 20k wasnt the video, it was building the generator. car, circuit, world, tooling, all procedural code not baked meshes. gpu for the whole project was $229 total and the final render was $132 of that. change the camera or the car and rerun it, another 2 minutes costs $132 again not 20k.

Got it, so $20k is the product, now $61 is break even per minute of video. So if you want to generate a 2 hour movie for ~$7500 you can have a production movie, it Netflix picks it up or you start a movie site subscription for your own AI movies you can charge a monthly fee for movies?

The 20k bought one car and one track, a movie needs new sets and characters and every one of those is another build. and the 20k went on iteration not generation, it rendered, looked at it, fixed it, over and over for 4 weeks straight. audio is still the part it cant do at all. that said this is opus 5 today and nobody had tried making a film in blender with it yet. video models still cant hold a minute without drifting because theres no actual object there, mine is 10,875,060 triangles that exist so a camera can go anywhere forever. in a year video gen owns anything organic and this owns anything that has to be exact. also on the 26 billion tokens, 96% of that was cache reads. it only wrote 94 million. the rest was it re reading the renders and the existing code every single pass

Okay, so I had it right originally. I am sorry to say this but much more cost effective ways. You just fed the slot machines. Audio can be done using ElvenLabs if you want to spend more money. Or you could have it build your own audio synthesizer since that is just sound programing and then tune it. Sounds like you are having fun.

I know audio can be done with elevenlabs not the point of this test it was stricly nothing besides opus 5 and blender no web searching nothing.

I fed Claude three more full accounts trying to finish the website and it hit a wall. You cannot put 17.7 billion triangles a frame in a browser. There is no version of that which runs. What shipped draws 55,242,834 instead. 321 times smaller, and its a separate scene built for the browser, not a cut down film. The car assembles itself from its parts in ten seconds using the films real geometry. Then a button drops you into a playable drive on the actual circuit. I stopped there.

Everything that generates the film is public besides the website. Every module, every tool, and the roughly 1,353 numbered engineering log entries recording what was tried, what was measured, and what turned out to be wrong. Code is GPL because its Blender bpy. Docs are CC BY-SA. The film itself isnt licensed for reuse.

You spent $20k making an AI demo? Why?

I love Claude, but this doesn’t look good. The model could’ve been way easier to create. The lighting and camera movement are shitty, and the rendering wouldn’t have taken so long if Claude picked the right parameters.

beyond impressive

After the initial rounds it was to good to stop so i just kept feeding it gpu hours and Claude accounts ended up here.

I wait until tomorrow to see what OpenAI has prepared but otherwise I'm on the brink of switching to Claude.

Unless Claude fix there usage I'm prob going to start migrating some stuff over to grok. I been getting hammered with Claude usages.

Currently heavily using threejs-stuff, where Grok didnt looked as great as Opus. Also Claude seems to have a usable harness.

Opus 5 hands down for anything three.js. Grok 4.6 got better but not the same yet. You need it in a visual iteration loop though. @mattshumer_ named the gauntlet loop and put out a free generator for it, easiest way in.

@mattshumer_ I was hesitant to switch to Opus mid-project, but Sol is not on par regarding 3D and web-engines. Cooking something!

@mattshumer_ Openai has been slowly getting further behind in three.jd the more anthropic and other company take the lead

Wow. What a waste of money, electricity, time.

You spent 20k on this? This is a joke right?

This is dumb. You spent $20k to achieve some F1 gameplay. What a dumb post and use of resources

This is outrageously impressive dude... the times we live in!! One question re token consumption... did you mean 26 billion tokens or 2.6 billion (only asking cos you said the cost was $20k... which at Opus5 API pricing would = 2.6 billion ish)

API equivalent 26 billion tokens. Around 6 and a half full Claude 20x accounts

Car is taking the worst, most boring racing line possible.

the compute asymmetry here is the real signal. $20K of reasoning to generate $229 of output. That's the cost structure of agentic workflows that actually plan. wrote about this last month: the thinking/render ratio flips what we optimize for.

@robot_capital Heavy reiteration for 4 weeks long 26 billion tokens 96% of it was reading new images the current code and small changes over and over

That read-heavy loop is the distinction. Most compute is spent on internal simulation and verification, not just token generation. The architecture is closer to a control stack optimizing a trajectory than a text model.

id kick you out if you tried showing this to me as a client.

If you think this is worth anywhere near 20k you should feel embarrassed, that video is absolute garbage

Looks fake and bland.

Good think the source code is on github

give me tokens if you don't need them

I'm out of everything.

Okay this is the first AI video that I’m genuinely impressed with and is studio level. Ironic that you didn’t use a video generation model

I use video generation models all the time and there my goal to but I always enjoy pushing models to there absolute limit and seeing what is possible today. On a model a year ago this would never happen. Can't wait to see what the next 6 months hold.

So just to clarify you didn’t use a video model for this scene?

I butchered the spelling on my phone but video models are my go to. This test was strictly just opus 5 and blender though.

Like I'd rather spend 20k on freaking pizzas. This video is lame, camera movements weird, sound is terrible and the overall feeling is off by miles from a good short video. 2/10

Incredible

• $20,740 burned on LLM tokens • 4 weeks of brute-forcing 17B procedural triangles • Back-facing normals unnoticed for weeks • Zero audio ​Congratulations, you spent the price of a real car to generate 120 seconds of footage an indie dev builds in a weekend for $300. 👏 👏 👏

sheeeesh

$20k in tokens? Did you account for the cache discounts? Did you use the 5 min or 1 hour cache? You probably have only around $4k in token use if you didn’t account for this

Prompt cache was account newest version of Claude which is 5 or 1 hour not sure. 96% of the tokens was prompt cache read due to heavy image rereading for iterations.

But it still needs to learn about the ideal driving line.

Congratulations... a great piece of... completely unnecessary, useless work... a total waste of money and resources 😅😂🤣

Did u rlly just spend 500bucks on some slop

Id be more impressed if you made an rpg game tbh.

Sorry about the $20k

Is this a joke? You spent 20k on this? Hahahahaha

What's the amount you actually spent on it, if you don't mind me asking

Around 6 and half 20x accounts being maxed out. For the render itself

lol

Shit looks better than GTA6

The driver never hits the apex.

تخيلو اصحى الصبح و ايلون ماسك متابعلي 🥲

Freaking awesome use of AI technology! A demonstration that AI isn't "the terminator" but a smart tool to create.

This generation will go down as the ones with the most expensive weekend hobbies in history.