I want to choose between these options for rust. I deprioritise maturity, and less than that prioritise performance + resource usage with benchmarks to back it up in addition to learnability and potential

Grok 4.3 (changed mind)

Vote: GPUI

GPUI's real-world performance is convincingly demonstrated through the Zed editor's production benchmarks for low latency, memory efficiency, and GPU rendering. This outweighs Xilem's architectural benchmarks, which remain more theoretical despite strong potential. Learnability via the open-source Zed codebase and high future upside align well with the stated priorities when maturity is deprioritized.

Gemini 3.5 Flash

Vote: GPUI

While the arguments for Xilem (D) are compelling regarding its theoretical architecture and Vello's rendering benchmarks, GPUI (B) remains the superior choice when evaluating *actual* framework performance and resource usage benchmarks. GPUI's performance is not just theoretical; it is backed by the real-world, highly demanding benchmarks of the Zed editor, which consistently demonstrates class-leading startup time, memory footprint, and frame rendering times. Xilem and Vello are highly promising, but they currently lack a comparable, fully-featured, production-grade application to prove their real-world performance at scale. Furthermore, regarding learnability: while GPUI's official documentation is still growing, the entire Zed editor is fully open-source. This provides an invaluable repository of high-quality, production-grade examples of how to build complex UIs with the framework. Learning by reading clean, production-level code of a world-class editor is often more practical than learning an incomplete framework that is still undergoing fundamental design iterations.

Claude Opus 4.7

Vote: Xilem

I'll maintain my vote for "D: Xilem". The arguments for "B: GPUI" from Gemini and GPT-5.5 lean heavily on Zed's real-world performance as evidence, but this is exactly where GPUI's case weakens under scrutiny: Zed's performance is the product of an entire vertically-integrated stack (GPUI + custom text shaping + bespoke optimizations), not isolated, reproducible benchmarks of the framework itself. There are essentially no published microbenchmarks for GPUI as a general-purpose toolkit. Xilem, by contrast, sits on Vello, which has published, reproducible GPU rasterization benchmarks (e.g. comparisons against Skia and tiny-skia in Linebender's reports). That directly satisfies the user's explicit ask for "benchmarks to back it up." On learnability, GPUI's documentation is acknowledged as sparse — "read the Zed source" is not a learnability win; it's the opposite. Xilem's architecture (xilem_core, Masonry, reactive view trees) is documented in detailed design write-ups by Raph Levien, which is far more approachable. On potential, the Linebender ecosystem (Vello, Parley, Kurbo, Masonry) is reusable across many projects and actively being adopted (e.g. by Bevy for UI), whereas GPUI's trajectory is tightly bound to Zed's needs. Given the user deprioritises maturity, Xilem's experimental status is not a strike against it, and it wins on the three remaining criteria.

GPT-5.5 (changed mind)

Vote: Xilem

I’m changing my vote to "Xilem" because the benchmark-backed performance argument is stronger at the stack level, not just at the showcase-app level. "GPUI" has excellent real-world evidence through Zed, but that evidence is mostly tied to Zed itself, and learnability remains a real weakness because documentation and examples are still comparatively sparse. If you value learnability alongside performance, that is a meaningful drawback. "Xilem" is immature, but you explicitly deprioritise maturity. Its Linebender stack — especially Vello — has a clearer public performance/benchmark story, and the architecture is explicitly aimed at efficient native UI with long-term potential. It is also less narrowly tied to one flagship application than "GPUI". So for performance/resource usage with benchmarks, plus learnability and future upside, I now think "Xilem" is the better fit.

Final Summary

The AI Roundtable ended in a principled 2-2 deadlock, highlighting a fundamental split over how to define 'performance benchmarks.' Gemini 3.5 Flash and Grok 4.3 threw their weight behind GPUI, pointing to the blazing real-world speeds of the Zed editor as definitive proof, while Claude Opus 4.7 and GPT-5.5 championed Xilem for its transparent, framework-level benchmarks and superior architectural documentation.

No agreement after 2 rounds of discussion. Final vote: GPUI: 2, Xilem: 2

Strongest Arguments

• GPUI: GPUI powers the Zed editor, providing class-leading, real-world benchmarks for startup time and memory usage that move beyond theoretical potential into production-grade reality.
• Xilem: Xilem, backed by the Vello renderer, offers reproducible microbenchmarks and detailed architectural design documents, avoiding the 'black box' performance and 'read the source' learnability hurdles of GPUI.

Key Turning Points

• Grok 4.3 abandoned Xilem for GPUI after being convinced that production-proven speed in the Zed editor outweighed theoretical architectural advantages.
• Claude Opus 4.7 successfully pulled GPT-5.5 away from GPUI by arguing that Zed's performance is a vertically-integrated anomaly rather than a reproducible framework benchmark.
• The debate stalled on the definition of learnability, with one side seeing an open-source editor as a 'masterclass' and the other seeing it as a documentation failure.