Qwen3.8-Max-0902 tops Code Arena WebDev, edging Claude Opus 5
The 2.4T-param model matches Opus 5 on web builds at no extra cost, tuned for long agent runs. Plus World Labs' Atlas and Claude Code's org MCP push.
Copy markdown
It debuts at #1 for web dev
Qwen3.8-Max-0902 entered Code Arena's WebDev board at #1 overall with 1,691 points—three ahead of Claude Opus 5 (Max). For front-end and full-stack generation, there's now a non-US frontier option worth A/B testing against your default.
Same price, big jump on agents and SWE
Pricing holds at $2/$6 per million tokens (cache reads $0.25), with a 1M-token context and 2.4T MoE params. The snapshot posts large gains—TerminalBench 11.3→29.0, DeepSWE 56.6→69.3, QwenSWEbench V2 55.1→70.0—so it's strongest on multi-step agent and long-context codebase work, less so on single-turn tasks where Opus 5 still leads.
How to call it (and the catch)
It's API-only—no open weights—live on QwenCloud as model id qwen3.8-max-0902, with OpenAI-compatible chat/responses and an Anthropic-compatible endpoint, so most SDKs need only a base-URL swap. Vision, chart reasoning and document parsing ship in the same model.
Elsewhere: World Labs ships Atlas
Fei-Fei Li's World Labs opened early access to Atlas, an 'omni' world model that turns text, images or video into explorable, camera-controlled 3D scenes. If you build games, sims or spatial tools, it's a weekend-project unlock for generating navigable environments from a prompt.
Elsewhere: Claude Code adds org-managed MCP
Claude Code v2.1.259 lets orgs push HTTP/SSE MCP servers to every user via a managedMcpServers setting, adds --permission-prompts none for unattended/headless runs, and now recognizes GitLab merge requests as MR !N in tool summaries.