GenUI Model Benchmarks
These benchmarks measure how fast each model responds and how often it produces structural hallucinations. What they can't measure is whether the interface a model builds is actually good for the person using it.
Learn how we benchmarked LLMs for GenUI.
Updated September 3, 2026
- 33 Model configurations, thinking and no-thinking counted apart
- 1,656 ms Fastest average round-trip, gemini-3.5-flash-lite-no-thinking
- 0.0% Structural hallucinations, highest rate of any model
- 3 × 5 Turns × iterations behind every average
Every model, timed.
Select a column to sort. Hover a header for its definition.
| gemini-3.5-flash-lite-no-thinking | 1,656 ms | 4,968 ms | 0.0% | 453.0 | 3 × 5 |
| gemini-3.5-flash-lite | 1,716 ms | 5,147 ms | 0.0% | 476.0 | 3 × 5 |
| gemini-3.1-flash-lite-no-thinking | 1,866 ms | 5,599 ms | 0.0% | 482.0 | 3 × 5 |
| gemini-3.1-flash-lite | 1,915 ms | 5,744 ms | 0.0% | 479.6 | 3 × 5 |
| gemini-3.6-flash-no-thinking | 2,908 ms | 8,724 ms | 0.0% | 525.4 | 3 × 5 |
| gpt-5.4-mini | 3,089 ms | 9,268 ms | 0.0% | 367.7 | 3 × 5 |
| gemini-2.5-flash-no-thinking | 3,267 ms | 9,801 ms | 0.0% | 584.9 | 3 × 5 |
| gpt-5.4-mini-no-thinking | 3,370 ms | 10,111 ms | 0.0% | 472.1 | 3 × 5 |
| gemini-3-flash-preview-no-thinking | 3,492 ms | 10,477 ms | 0.0% | 565.1 | 3 × 5 |
| gemini-2.5-flash | 3,602 ms | 10,806 ms | 0.0% | 535.7 | 3 × 5 |
| gpt-5.4 | 3,678 ms | 11,033 ms | 0.0% | 489.7 | 3 × 5 |
| gpt-5.6-luna | 3,721 ms | 11,163 ms | 0.0% | 585.2 | 3 × 5 |
| gpt-5.6-luna-no-thinking | 3,835 ms | 11,506 ms | 0.0% | 584.9 | 3 × 5 |
| gemini-3.7-flash | 3,840 ms | 11,519 ms | 0.0% | 553.0 | 3 × 5 |
| kimi-k2.7-code-highspeed | 4,090 ms | 12,269 ms | 0.0% | 610.4 | 3 × 5 |
| gpt-5.4-no-thinking | 4,407 ms | 13,222 ms | 0.0% | 541.7 | 3 × 5 |
| gpt-5.6-terra-no-thinking | 4,447 ms | 13,341 ms | 0.0% | 534.8 | 3 × 5 |
| gpt-5.6-terra | 4,499 ms | 13,497 ms | 0.0% | 617.7 | 3 × 5 |
| gemini-3.5-flash-no-thinking | 4,526 ms | 13,578 ms | 0.0% | 584.5 | 3 × 5 |
| gemini-3.6-flash | 4,979 ms | 14,938 ms | 0.0% | 534.3 | 3 × 5 |
| deepseek-v4-flash-no-thinking | 5,261 ms | 15,783 ms | 0.0% | 556.7 | 3 × 5 |
| claude-haiku-4-5 | 5,279 ms | 15,837 ms | 0.0% | 518.5 | 3 × 5 |
| gemini-3.8-flash | 5,604 ms | 16,811 ms | 0.0% | 561.4 | 3 × 5 |
| claude-sonnet-5 | 5,665 ms | 16,995 ms | 0.0% | 561.9 | 3 × 5 |
| deepseek-v4-flash | 5,971 ms | 17,913 ms | 0.0% | 612.1 | 3 × 5 |
| gemini-3-flash-preview | 6,322 ms | 18,967 ms | 0.0% | 568.3 | 3 × 5 |
| gpt-5.5-no-thinking | 6,977 ms | 20,931 ms | 0.0% | 530.8 | 3 × 5 |
| gemini-3.5-flash | 7,788 ms | 23,365 ms | 0.0% | 538.0 | 3 × 5 |
| claude-opus-4-8 | 7,818 ms | 23,453 ms | 0.0% | 616.3 | 3 × 5 |
| gpt-5.5 | 7,976 ms | 23,929 ms | 0.0% | 608.5 | 3 × 5 |
| claude-sonnet-4-6 | 8,193 ms | 24,578 ms | 0.0% | 553.4 | 3 × 5 |
| mercury-2 | 8,375 ms | 25,124 ms | 0.0% | 545.2 | 3 × 5 |
| kimi-k2.7-code | 17,056 ms | 51,167 ms | 0.0% | 666.5 | 3 × 5 |
Token stats are shown only when the provider reports them.
Column glossary.
- Model
- The model id. A -no-thinking row is the same model with thinking disabled. Rows are sorted fastest-first.
- Avg round-trip
- Mean wall-clock time per turn, from request sent to stream complete, over successful turns. Comparable across providers; this is the headline metric.
- Avg per pass
- Mean time to complete one full pass (all turns in an iteration summed). Comparable across providers.
- Structural hallucinations
- Share of turns with a structural hallucination — a hallucinated catalog item or prop, or output that didn't conform to the catalog schemas or the A2UI spec — plus turns that errored outright or produced no renderable surface. Hover the value for the breakdown.
- Avg output tokens
- Mean completion tokens per turn (excludes the prompt). Includes reasoning tokens for thinking models, so it drops on -no-thinking rows.
- Turns × iters
- Turns per iteration × iteration count behind each average.
Reading the table.
How the numbers were taken, and what they can’t tell you.
01 Which number should we compare first?
Avg round-trip. It is the headline metric, the mean wall-clock time per turn from request sent to stream complete, and the table opens sorted on it, fastest first.
02 How many runs sit behind each average?
The Turns × iters column shows it row by row. In the run dated at the top of the page, every model had 3 turns per iteration over 5 iterations.
03 Does a failed request count against a model?
Yes. The structural hallucination rate also counts turns that errored outright or produced no renderable surface. Hover a value above zero for the breakdown.
04 What does a -no-thinking row mean?
It is the same model with thinking disabled. Read each pair side by side to see what thinking changes in round-trip time and output tokens.
05 Can we compare token counts across providers?
With care. Token stats appear only when the provider reports them, and for thinking models they include reasoning tokens. The two timing columns are the ones the page calls comparable across providers.
06 Does the fastest model build the best interface?
The table can’t tell you. It times each model and counts structural hallucinations, and whether the interface a model builds is good for the person using it is something these benchmarks can’t measure.
We’ll care about your success as much as you do.
Our unicorns read every message and reply, usually the same day.
Start the conversationA 30-minute call about what you’re building.






