You may think the frontier models structure their thinking somewhat similarly. Our research found that their generation process differ wildly, and you can see it in the output. Watch as we compare @OpenAI's GPT-6 Astra and @Anthropic's Claude Opus 5.5 https://x.com/i/article/2106136192739897344
Builder vs. Designer: Claude Opus 5.5 vs. GPT-6 Astra
Give Claude Opus 5.5 and GPT-6 Astra the same prompt to build a game, and they can return with surprisingly different ideas of what ‘finished’ means. In a typical Design Arena game, Opus finished in under five minutes; Astra took closer to eleven. The difference went beyond speed: each model spent its effort on a different part of the experience.
The Race
Across about 400 games (241 from Opus 5.5, 166 from Astra), Opus 5.5 produced code roughly 1.3 times faster than Astra, measured in characters per second of total generation time. Its output was also much leaner: a median of around 40,000 characters, compared with about 70,000 for Astra. That's roughly 44% less code per task.
Opus finishes sooner because it produces code faster and writes less of it. To understand where that difference comes from, we compared the models’ thinking summaries with their finished games. Two tendencies emerged: Opus behaves like a builder, and Astra like a designer.
Those priorities show up in everything from ball physics to page layout.
The Builder
Opus’s summaries focus on a practical question: how will the game behave when someone plays it?
Opus commits early. In about four of every five summaries, Opus weighs different approaches and then picks one decisively. Astra shows that pattern in just over a quarter. In a free-kick game called Touchline 3, for example, Opus decided early that the player's aim would account for the ball's curve, while wind would be left for the player to judge. That decision held all the way through to the final code.
Next comes the math. Both models include about a dozen numbers per summary, but Opus’s numbers tend to describe motion. In Touchline 3 it worked out how hard the ball needs to be kicked to land on target, deriving the upward speed from flight time and gravity. The shipped code uses the same calculation.
Opus also drafts code inside its summaries: ball physics, goalkeeper dives, the net. This happens in 91% of its summaries, mostly in the second half. Astra does it in about one in five, usually near the end.
Opus also checks its plan for playability. Both models check their work about equally often, but Opus's checks cluster near the end and focus on gameplay, like whether a well-struck shot can actually clear the wall and a jumping defender. It also plans small kindnesses for the player, such as a moment of safety after taking a hit or a little grace time to jump after running off a ledge. That shows up in 27% of its summaries, versus 8% for Astra.
The page around the game gets much less attention. Only 16% of its games ship with a header, footer, or navigation, and Touchline 3 has none at all.
Astra puts more of its effort into that surrounding experience.
The Designer
Astra often starts with the page: what surrounds the game, how it is presented, and what a visitor needs to get started.
In Dust and Lead, a Western-themed shooting gallery, Astra's very first thoughts were about the page: the header, roomy spacing, a footer holding the controls, and a cream-and-dark color palette. All of that came before a single rule of the game. Three quarters of Astra's summaries discuss the surrounding page, compared with just over a quarter for Opus. And it follows through: 92% of Astra's games ship with a header, footer, or navigation, against 16% for Opus. In time terms, Astra spends about 22 seconds of each game's thinking on the page experience; Opus spends about 2.
Astra’s planning is also more structured. Astra often sets itself a budget the prompt never asked for, such as noting roughly 30,000 tokens available and aiming for about 10,000. It still ends up writing longer games than Opus on average. Next it maps out the layout by screen width: in Dust and Lead, a sheriff's station on the left, a saloon in the center, and a bank on the right, each at a fixed fraction. Its summaries even use sequential stage headings, moving from planning to designing to evaluating to refining. The actual game rules don't show up until about a quarter of the way in.
Astra’s summaries use more tentative language. Even the opening screen is described as having perhaps one or two bandits on the right. Hedging words like "maybe", "might", or "it seems" appear in nearly every Astra summary, roughly 20 times as often as in Opus's.
Finally, Astra finishes out the product with help screens, pause functions, and accessibility features. Every Astra game in our code sample included accessibility-related code, such as screen-reader labels or reduced-motion handling. Opus did so in about 9% of its games. In 60 Astra games, accessibility appeared in the code without being mentioned in the summary. It seems to be a recurring implementation pattern, even when the summary doesn’t call it out.
Why Not Both?
A successful product benefits from both priorities, and each model shows traces of the other’s style. Dust and Lead plans forgiving hitboxes so shots are easier to land, which is a very builder-like thought. And Opus plans help screens about as often as Astra does, even if it ends up building fewer of them.
The difference in emphasis helps explain the gap in output size. Astra's games run about 27,000 characters longer on average. Almost half of that extra is styling, about a quarter is the page itself, and roughly a tenth is menus, help, and buttons. The code for the actual game is, if anything, slightly smaller in Astra's games. Much of the gap comes down to scope: Astra builds more around the game.
How Claude Got Here
Opus 5.5’s performance follows an uneven run for Claude. Measured while each version was the newest Claude on the arena, the share of game votes won went from 66% with Opus 4.6 to 61% with Opus 4.7, then dropped to 47% with Opus 4.8, before recovering to 61% with Opus 5 and climbing to 75% with Opus 5.5.
Opus 4.8 was a clear step backward. Its games shrank to less than half the size of Opus 4.6's, and they lost much of their life. 35% of Opus 4.6's games had something moving on the opening screen; for 4.8, that fell to 5%. When we read 12 of its games closely, three didn't run at all and five had a control or stat that did nothing: a power meter the shot ignored, a health bar that never moved, a play button that only showed a fake loading screen.
From Opus 5 onward, the games improved noticeably, especially in how they sound and feel. Opus 5.5 writes about the same amount of code as its predecessor, but faster, and its games are livelier still: 43% move on their own from the first screen. In one boxing game, the title screen runs the game's real punch code with an orbiting camera and a crowd of 560 spectators. The improvement shows up in sound, feel, and fairness, without requiring a larger output.
Which One?
We started with a race. The more useful question turned out to be what each model considered worth building. Opus concentrated on the game: the curve of a free kick, the height of the defensive wall, how far the keeper could reach. Astra invested more in the surrounding experience: layout, controls, screen-reader support, and pausing when you switch tabs.
For these game-building tasks, Opus tended to prioritize playable mechanics and finish sooner. Astra tended to deliver a more elaborate product around those mechanics. The right choice depends on what you want to build. Give both the same idea and compare where they put their effort—that’s what Design Arena is built to do.
Try it yourself at designarena.ai. And if you'd like more breakdowns like this, follow us on X, LinkedIn, Discord, and Instagram.
Written by Alex Gong (@gongzhongcheng), engineer @Intelligence_ai