# Qwen 3.8 vs. Claude Opus 5: Which Model Builds Better Software? Benchmark scores are a useful shorthand, but they don't answer the question developers actually face: which model produces better working code when given a hard, open-ended task? **Qwen 3.8 scores 58 on the Artificial Analysis Intelligence Index. Claude Opus 5 scores 63. A five-point gap looks like a clear verdict. It isn't**. This article covers what both models are, how they differ in architecture and cost, and what happens when you give each of them the same prompt: build a complete playable 3D endless runner game in a single HTML file. ## The models ### Claude Opus 5 Claude Opus 5 is Anthropic's flagship proprietary model, sitting at the top of the Artificial Analysis Intelligence Index with a score of 63. ![The Artificial Analysis Intelligence Index chart, highlighting Qwen's score of 58, which is on par with many top-tier proprietary models.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/c2439da5-2a59-493e-9720-71a4415cf300/md2x =1920x1080) It's closed-source with no public weights. The only access path is through Anthropic's API. It's known for strong reasoning, large context handling, and polished output. It's also faster than Qwen and generally produces more immediately usable results for interactive tasks. ### Qwen 3.8 (2.4T-A95B) Released by Alibaba on August 12, 2026, Qwen 3.8 2.4T-A95B is the first time Alibaba has released open weights for a Qwen-Max-class model. It scores 58 on the same Intelligence Index, the same score as its proprietary counterpart Qwen 3.8 Max. The architecture is a sparse MoE with 2.4 trillion total parameters and 95 billion active parameters per token. It routes each token to 10 specialized expert sub-networks out of 512, plus one shared expert, keeping per-step compute much lower than the total parameter count suggests. ![A visualization of Qwen's Mixture of Experts (MoE) architecture, showing 2.4T total parameters with 95B active per token.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/d53f7d3f-fc80-4fc4-5e28-1d9b4645ca00/md2x =1920x1080) The native context window is 262,144 tokens, extensible to approximately 1 million. The open-weight checkpoint is text-only and always reasons: every response begins with a think block and thinking mode cannot be disabled. The weights ship under a custom Qwen3.8-Max License, not Apache 2.0. It's free for most commercial use, but products above 100M monthly active users or $20M monthly revenue have additional display requirements. Read the license file before building on it. Self-hosting requires serious infrastructure: at minimum two NVIDIA B300 or AMD MI355X nodes. Most developers use a hosted inference provider. Pricing runs around $2 per million input tokens and $6 per million output tokens. Throughput is roughly 45–48 tokens per second. ### Key differences at a glance | | Qwen 3.8 | Claude Opus 5 | | :--- | :--- | :--- | | Intelligence Index | 58 | 63 | | Weights | Open : Qwen3.8-Max License | Closed | | Context (native) | 262K, ~1M extended | 200K | | Cost (hosted) | ~$2/$6 per 1M tokens | ~$5/$25 per 1M tokens | | Speed | ~45–48 tok/s | Faster | | Thinking | Always on, open weights | Optional | ## The test: a 3D endless runner in one HTML file Both models received the same prompt with no follow-up corrections. Ten specific requirements, all measurable: ![A clear, stylized list of the 10 build requirements for the 3D endless runner game.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/187fe6e7-99c4-409f-9dde-412c7fd52000/md2x =1920x1080) Three.js from a CDN; automatic forward movement; lane switching between left, center, and right; jump and slide mechanics; procedurally generated path with turns, gaps, and obstacles; score based on distance survived; game over and restart; smooth follow camera; clean and well-structured code; fun and playable for 30–60 seconds. The single HTML file constraint with no external assets forces every visual element — player, obstacles, environment — to be generated through code using Three.js primitives. ## How each model approached the task ![A side-by-side terminal view showing the identical prompt being entered for Qwen (left) and Claude Opus 5 (right).](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/9fee23ec-8d9a-420b-6584-25ea782b3400/lg2x =1920x1080) Claude started generating code immediately, scaffolding the HTML document and setting up the Three.js scene, camera, and renderer in rapid succession. It built components piece by piece: the 3D world, the player, the path, then game logic. The process looked fast and confident. Qwen produced no code for a significant portion of its generation time. It wrote an architectural plan first: game loop logic, player state management, collision detection rules, procedural generation strategy, camera movement, UI for score and game over. Only after this planning phase did it begin generating implementation. This difference in approach is the core of what the test reveals. Qwen's slow, verbose planning phase looks like a weakness. In this case, it was the opposite. ## The results ![A split-screen comparison showing the final, playable games generated by Qwen (left) and Claude Opus 5 (right).](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/a1874b68-be3b-4372-92ce-d19a3f0caa00/orig =1920x1080) ### Claude Opus 5 The game generated by Claude was visually polished. It featured a character running through a temple-like corridor with pillars and torches, coin collection mechanics, lane switching prompts, and a styled UI for score, distance, and speed. It looked like a real game. All ten requirements were met. The controls were responsive, the camera followed smoothly, the game over screen worked, and the restart function was clean. **Claude's game was fully playable and genuinely fun.** It was the clearer winner for this specific task. ### Qwen 3.8 Qwen's game was visually basic: a block for the player, an outdoor road, rudimentary trees and obstacles. The aesthetic was rough. More critically, the procedural generation produced impossible sequences in places, gaps immediately followed by walls, making the experience feel unfair. The camera was less smooth, and the overall feel was an early-stage prototype. It technically met most requirements but fell short on the most important one: it wasn't consistently playable or fun. For this task, Claude produced the better product. ## What this actually shows The result might seem to confirm the benchmark gap. But the more interesting finding is what drove each outcome. Claude's speed and directness produced polished results when the task was well-defined and had clear aesthetics. It translated a prompt into something that looked and felt finished. For game development, where feel and visual cohesion matter, that directness was an advantage. Qwen's planning-first approach didn't pay off here. But consider what happens when the task is harder to define, involves more interconnected systems, or requires careful decisions before committing to code. The same verbal planning behavior that produced a clunky game would be a meaningful asset for an agentic workflow building out a full application backend or refactoring a large codebase. The deliberate architecture-before-code style is better suited to tasks where premature implementation compounds into hard-to-fix problems. ### Cost at scale This is where the comparison changes significantly for developers building production systems. ![The Cost per Intelligence Index Task chart, showing Qwen 3.8 is significantly cheaper than Claude Opus 5 for a standardized task.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/069e6478-80cf-4e37-9d93-53d6ecc2a800/lg2x =1920x1080) Qwen at $2/$6 per million tokens versus Claude Opus 5 at $5/$25 makes a substantial difference when a model is called hundreds or thousands of times in an agentic loop. A task that costs $0.50 per run with Claude costs roughly $0.10 with Qwen. At 10,000 runs, that's $4,000 versus $1,000. For high-volume workflows, the intelligence gap may be worth less than the cost gap. ### Open weights The ability to self-host Qwen is a real option for teams with the infrastructure, even if the 2.4T model itself requires data-center hardware. The Qwen family also includes much smaller models (a 27B version is available under Apache 2.0) that retain meaningful capability while being realistic to run on private infrastructure. This flexibility doesn't exist with Claude. ## When to use each **Claude Opus 5** is the better choice for: interactive coding sessions where quality and speed matter; tasks where the output needs to feel polished on the first pass; one-off creative coding challenges where visual cohesion is part of the requirement. **Qwen 3.8** is the better choice for: agentic workflows making many API calls where cost compounds; tasks requiring large context windows; teams needing open-weight flexibility for self-hosting or fine-tuning; long-horizon planning tasks where verbose architectural reasoning is an asset rather than friction. The benchmark gap is real. Claude Opus 5 is more capable on the tasks those benchmarks measure, and for this specific game-building challenge, it showed. But the choice between these models should be driven by the nature of the work and the cost constraints of the project, not just the leaderboard position.