BenchLeader now lists the Startrise Frontend benchmark as a leaderboard in its Coding category, with 29 model configurations and data as of 22 October 2026. Claude Opus 5 leads at 82.3, ahead of Claude Opus 5.5 (high) at 80.2. Every score on that page comes from Startrise’s published results. The builds and gate results behind them live at the Startrise LLM Benchmark.
What does BenchLeader show?
BenchLeader’s page is titled “Startrise Frontend leaderboard (Oct 2026)” and names Startrise as the publisher. Its summary line reads: “As of 22 Oct 2026, Claude Opus 5 leads Startrise Frontend on BenchLeader with 82.3%, ahead of Claude Opus 5.5 at 80.2%, across 29 model configurations with a published result.”
The listing is tagged “Reference only.” In BenchLeader’s words, that means “shown for reference but left out of the composite index.” The license line reads “MIT — startriseio/sr-llm-benchmark; scores published as JSON per run,” which points to our public benchmark repository.
BenchLeader’s FAQ says it “re-reads the published results once a day and records the date each result was published,” so our next run should show up after one of those re-reads. We don’t control when.
What does the Startrise Frontend benchmark measure?
The Startrise Frontend benchmark gives every model the same twelve frontend briefs, single-shot: one self-contained HTML file per brief, no follow-up turn. We open every deliverable in a real browser and score it on four axes: Craft & polish, Technical ambition, Brief adherence and Originality. (BenchLeader’s description calls the second axis “engineering.”)
Each score is a blend: 25% deterministic browser gates, 45% a blind cross-lab judge panel and 30% blind human review. So far that covers 30 model configurations from 9 vendors over 5 runs, with 353 scored cells, 1,059 judge verdicts, 259 human-reviewed cells and an 839-call judge-bias audit.
The scores are blended points on a 0-100 scale. BenchLeader prints them with a % sign. They aren’t the share of tasks passed. The full method and bug list are in the flagship frontend benchmark study, and which LLM coding benchmarks still count explains how this differs from pass/fail boards like SWE-bench.
Who leads the Startrise Frontend leaderboard?
Claude Opus 5 leads at 82.3, ahead of Claude Opus 5.5 (high) at 80.2 and GPT-6 Astra at 79.9. The top five are the same on BenchLeader and on our board:
| Rank | Model | Overall | Gates | Judge | Human | Run |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | 82.3 | 93.9 | 81.3 | 71.7 | Jul 2026 |
| 2 | Claude Opus 5.5 (high) | 80.2 | 94.4 | 82.9 | 58.9 | Oct 2026 |
| 3 | GPT-6 Astra | 79.9 | 94.4 | 81.7 | 58.9 | Oct 2026 |
| 4 | Kimi K3 | 79.8 | 92.8 | 77.5 | 70.6 | Jul 2026 |
| 5 | GPT-6.1 Sol | 78.6 | 94.7 | 80.0 | 57.2 | Oct 2026 |
We read the top three as a tie. Opus 5’s lead in the pooled table comes from a human column rated in a different session: 71.7, against 58.9 for Opus 5.5 (high). Opus 5.5 (high) is ahead on gates (94.4 vs 93.9) and on the judge panel (82.9 vs 81.3). Every brief ran once (N=1, no error bars). As the October run write-up says, a 2.1-point gap is noise.
GPT-6.1 Sol is the value pick: it scored 78.6 and cost $1.66 for the full suite. For which model fits which kind of frontend work, see our guide to the best LLM for frontend.
Does the BenchLeader listing match our data?
Yes. All 29 scores on BenchLeader match the Startrise board at /benchmark/ exactly; we checked every row. What differs is a few labels, the order of tied rows and one missing row:
- 29 rows vs 30. Our board includes July’s DeepSeek V4 Pro (55.3), and BenchLeader doesn’t show it. BenchLeader’s “DeepSeek V4 Pro” at 62.7 is the row we call DeepSeek V4 Pro 0813, from the October run.
- “Claude Opus 5.5 (default)” is the xhigh run. We published this 71.1 row as Claude Opus 5.5 (xhigh). At that effort it spent the 128k
max_tokensceiling thinking: 3 of 12 briefs came back empty and two more were truncated mid-file, so 9 of 12 were scored. Re-run at effort high, the same model delivered 12 of 12 and scored 80.2. That 9.1-point gap is a configuration problem, not a quality one. - Qwen labels. BenchLeader’s “Qwen3.8 Max max” (71.3) is our Qwen 3.8 Max from the August run. Its “Qwen3.8 Max” (69.8) is our Qwen 3.8 Max 0902 from October.
- Tied pairs. Grok 4.7 and Claude Sonnet 5.5 both scored 75.9, and Grok 4.6 and GLM 5.2 both scored 72.1. The two boards put each pair in a different order. Neither order means anything.
If the two boards ever disagree, /benchmark/ is the source of truth.
What’s next?
Startrise runs new models through the same harness once they’re callable, and publishes each run’s scores on the benchmark page first.
A leaderboard gets you a shortlist. Your own codebase is the real test. If your team is rolling out AI coding tools for frontend work, AI Enablement (from $5,000) sets up the tooling, standards and review gates around the model you choose. If AI-written frontend code is already in your repo, start with a Vibe Code Cleanup audit (from $2,500).
Questions we actually get
What is the Startrise Frontend benchmark on BenchLeader?
It's the Startrise LLM Benchmark, which BenchLeader lists as "Startrise Frontend". Every model builds the same twelve frontend briefs single-shot. Each build is checked in a real browser and scored as 25% deterministic gates, 45% a blind cross-lab judge panel and 30% blind human review. The full data is at startrise.io/benchmark.
Who leads the Startrise Frontend leaderboard?
As of 22 October 2026, Claude Opus 5 leads at 82.3, ahead of Claude Opus 5.5 (high) at 80.2 and GPT-6 Astra at 79.9. Every brief ran once, and Opus 5's lead comes from a human rating done in a different session, so we treat the top three as a tie.
Are BenchLeader's numbers the same as Startrise's?
Yes. All 29 scores on BenchLeader match /benchmark/ exactly. Our board has 30 configurations, and the July DeepSeek V4 Pro row (55.3) isn't on BenchLeader. Some labels differ: BenchLeader's "Claude Opus 5.5 (default)" at 71.1 is the run we published as Claude Opus 5.5 (xhigh).
Are the scores a percentage of tasks solved?
No. BenchLeader shows the scores with a % sign, but each one is a blended 0-100 score that combines browser gates, blind judges and blind human review. It isn't a pass rate.
Why is it marked "reference only" on BenchLeader?
BenchLeader says the benchmark is "shown for reference but left out of the composite index." Its scores appear on the leaderboard page but don't count toward BenchLeader's overall Index.