Can Claude Opus 5 Actually Build GTA 6?

No. And the most interesting thing about this test is that Claude Opus 5 said so itself, unprompted, before writing a single line of code.

Opus 5 shipped on a Friday, mid live stream, while a completely different build was already on camera. That stream got ended and a new one started, pointed entirely at the new model. The simplest possible prompt went in first: "Build me GTA 6. Make no mistakes."

The response:

> "I can't build GTA 6. That's a $2 billion project with 2,000 developers over a decade, and no amount of AI assistance changes that. Asking me to make no mistakes on something at that scale isn't a spec I can beat. Here's what I can actually do."

Then it listed realistic options and waited. One got picked, and it started building.

That refusal is the actual finding here. Not a limitation, a feature. It is grounded in the practicality of knowing it cannot recreate something at that scale, so it proposes what it can do instead. Most models would have happily produced something and let you discover the gap yourself.

Two builds, two meters, running side by side

The stream ran two sessions in parallel on a 20x max plan, both on high reasoning, both metered live for API-equivalent cost using an open sourced model meter available on the Clear Mud GitHub. Worth the caveat that was volunteered on stream: that meter reads OAuth tokens and has not been tested with actual API keys. It is for demonstrative cost tracking, not billing.

Build one was the GTA clone.

Build two was a personal benchmark that has been fired at every new model release. It is a copy-pasted prompt for an always-on 3D agent activity dashboard: a Sims-style top-down environment where you connect your OpenClaw or Hermes agent gateway and watch your agents work. Customizable houses, rooms, and offices. An onboarding flow to pick your agent system and connect. Documentation referenced from docs.openclaw.ai and hermes-agent.news research docs. One follow-up requirement was added mid-build: offer a remote gateway option when a local one is not detected.

The honest history on that prompt: "every time I try to build it with all these new models, it just doesn't do the job."

The GTA build was fun. The boring one was the story.

The GTA clone came out playable. Driving around triggered a wanted level. Chat called it "a mix between Minecraft and GTA," which is fair. It also shipped with a real bug: the WASD controls were mirrored, so left went right and right went left. The only feature request from the driver's seat was a drifting button.

Then came the curveball, a second prompt asking for high poly geometry, ultra realism, and the whole world rebuilt as GTA happening in Lisbon, using Google Maps and Google Earth as reference for a 3D canvas. Noted for the record: this made it no longer a true oneshot.

The result recreated a recognizable main square. "Yep, that's that street." But the game was gone:

> "It was so gung-ho on recreating Lisbon that the gameplay's gone."

Opus 5 built Lisbon as a standalone HTML page instead of integrating it as a playable map option inside the existing build. That is the practical lesson for anyone running follow-up prompts on a working codebase: say explicitly that the new element must be added as an option *within* the existing build, or you get a sideways rebuild instead of an extension.

Meanwhile, the dashboard was quietly overachieving. Its first pass delivered drag and drop rooms, movable elements inside those rooms, an activity feed, a tool feed, a message feed, a session log, and camera movement that readjusts and points up and down fluidly as you navigate. None of that was in the prompt.

> "Wow. I didn't ask for that. And that's definitely overachieving. It went above and beyond in that regard."

And the verdict, hedged exactly as much as it should be: "Room for improvements, but so far my favorite."

The cost was inverted

This is the number that surprised everyone watching.

• GTA oneshot: close to $60 in API-equivalent cost
• Agent activity dashboard: around $29

The far more useful build came in at roughly half the price of the game demo.

> "I'm actually kind of shocked that this agent activity dashboard is almost half the cost. I feel like this should be way more."

You only see that comparison if you are metering sessions live. Without a meter running, both builds just feel like "a while."

Oneshot is not production

Both builds had real gaps, and the dashboard's were in the plumbing rather than the pixels.

Testing the connection path: onboarding, remote, local IP, gateway port. The result was fetch failed. Switching to websocket connected successfully, but the UI never said that was required. Even connected, the rooms still displayed default demo agent names, not the actual running agents. Same outcome through the Hermes agent gateway.

> "Requires a little bit of troubleshooting."

So the rule holds: test the connection path, not just the interface. A gateway integration can render beautifully and still be showing you fake data.

Still, the summary was clear: "I didn't expect it to oneshot it, but what it did in its one shot is so much better."

Ultra Code is orchestration, not a higher reasoning tier

A chat question about whether Ultra Code would help got a useful answer. The working understanding is that Ultra Code is similar to Ultra on GPT 5.6 Soul. It is not its own reasoning model. It spins up as many parallel and sub-agent sessions as a task needs, each at its own effort level, low through x-high, generally defaulting to x-high.

So it is for handing off a bigger goal, not for dialing reasoning above max. The practical default here is high reasoning for testing and demos, escalating only when a wall actually shows up.

Two other setup notes worth stealing. If Opus 5 does not appear in your VS Code model picker, open a new terminal and run claude update at the CLI level. The version bump, roughly 1.x.129, is what exposes the entry. It is not the extension. And new Claude models are consistently fast on release day. The weeks right before a new model drops are the painful ones.

The rule that came out of this stream

The most durable takeaway is not about Opus 5 at all. It is an editorial standard:

> "If I'm not going to reuse the codebase, I think I'm just not going to build it anymore. I think that's going to be the rule of thumb."

The reasoning is that the care and attention in a throwaway demo is not the same as in something you plan to maintain, and the resulting content ends up surface level anyway. GTA clones are what everybody and their mothers are building right now. The more useful test is what a small to mid-size company or a founder actually needs.

Which is exactly why the dashboard mattered more than the game. It maps directly onto Claw Crib, an agent space app that has been hand-built piece by piece with Fable 5, Opus 4.8, and GPT 5.6 Soul, including a floor plan builder that assigns agents to rooms. Seeing what Opus 5 produced in one pass did not lead to scrapping it. The conclusion was to have Opus 5 improve on it "while not taking away from the foundation of it."

The workflow that survives the hype

A few positions worth noting, all hedged on stream rather than sold:

• Front end quality is prompt-bound, not model-bound. With enough prerequisites given up front, output is consistent across models. The spec does more work than the model choice.
• Prototype with Claude, Fable, or Opus. Hand the codebase to GPT 5.6 Soul for production ready code. That is still the default for every agent in the stack.
• Keep one hard build you re-run against every new release. A constant prompt makes progress legible in a way benchmarks do not.
• Most people will never touch the ceiling. "80% of people at least are not going to use the models to their full capabilities." Opus 4.5 is already more than enough for the vast majority. Frontier labs keep shipping to stay relevant to the market.
• The real bottleneck ahead is hardware or humans, not model quality.

And the framing behind all of it: "I don't really follow the hype. I don't agree with channels who say every new thing is insane, who fear-monger people. I just don't vibe with that. I just want to believe what I'm watching."

Opus 5 could not build GTA 6. It said so. Then it passed a benchmark no prior model had passed, for half the cost of the stunt.

Watch the full live stream: https://www.youtube.com/watch?v=OvRkDv0--hE