Web Clone Bench v1 (Web Clone Bench)
We show this table for reference; we do not rank on it.
Coding agents recreate nine reference web apps from an empty directory. Plato checks visual and functional behavior through end-to-end workflows.
Full score (%) on Web Clone Bench — October 15, 2026 snapshot
We mirror the published full score (%) view for Web Clone Bench. Claude Opus 5.5 (xhigh) leads the public snapshot at 34.2%, followed by Claude Opus 5.5 (max) (32.3%) and Claude Opus 5.5 (high) (30.5%). We do not use these results to rank models overall.
Claude Opus 5.5 (xhigh)
Anthropic
Vendor coding harness (source description) · xhigh reasoning
Claude Opus 5.5 (max)
Anthropic
Vendor coding harness (source description) · max reasoning
Claude Opus 5.5 (high)
Anthropic
Vendor coding harness (source description) · high reasoning
29 effort configurationsCodingCurrentDisplay onlyUpdated October 15, 2026 snapshot
| Rank | Model / effort | Full score | Avg cost / run | Avg time / run | Steps | Parameters (B) | Open / closed |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5xhigh effort | 34.2% | $652.07 | 5h 36m | 10,305 | Not reported | Closed |
| 2 | Claude Opus 5.5max effort | 32.3% | $718.71 | 5h 46m | 10,571 | Not reported | Closed |
| 3 | Claude Opus 5.5high effort | 30.5% | $546.59 | 5h 57m | 10,302 | Not reported | Closed |
| 4 | Claude Opus 5.5medium effort | 18.6% | $403.88 | 4h 4m | 6,609 | Not reported | Closed |
| 5 | GPT-6 Astramax effort | 17.5% | $207.27 | 6h 20m | 1,290 | Not reported | Closed |
| 6 | Claude Fable 5.1high effort | 13.3% | $272.24 | 2h 13m | 2,768 | Not reported | Closed |
| 7 | Claude Fable 5.1xhigh effort | 13.0% | $314.08 | 2h 32m | 2,807 | Not reported | Closed |
| 8 | GPT-6.1 Solmax effort | 12.6% | $25.35 | 6h 40m | 793 | Not reported | Closed |
| 9 | GPT-6 Astraxhigh effort | 11.8% | $116.64 | 2h 39m | 585 | Not reported | Closed |
| 10 | Claude Fable 5.1max effort | 11.6% | $227.84 | 1h 51m | 2,565 | Not reported | Closed |
| 11 | GPT-6 Astrahigh effort | 10.0% | $75.24 | 1h 42m | 434 | Not reported | Closed |
| 12 | GPT-6.1 Solhigh effort | 9.9% | $14.25 | 2h 45m | 628 | Not reported | Closed |
| 13 | GPT-6 Astramedium effort | 9.6% | $45.81 | 57m | 278 | Not reported | Closed |
| 14 | Claude Fable 5.1low effort | 9.2% | $174.75 | 2h 7m | 2,311 | Not reported | Closed |
| 15 | Claude Fable 5.1medium effort | 9.1% | $206.55 | 2h 9m | 2,188 | Not reported | Closed |
| 16 | GPT-6.1 Solxhigh effort | 8.6% | $16.78 | 3h 15m | 672 | Not reported | Closed |
| 17 | GPT-6 Astralow effort | 7.8% | $30.69 | 35m | 189 | Not reported | Closed |
| 18 | GPT-6.1 Solmedium effort | 6.9% | $5.80 | 1h 2m | 271 | Not reported | Closed |
| 19 | GPT-6.1 Sollow effort | 4.6% | $2.49 | 41m | 209 | Not reported | Closed |
| 20 | Grok 4.7low effort | 2.8% | $18.32 | 26m | 532 | Not reported | Closed |
| 21 | Grok 4.7high effort | 2.7% | $19.22 | 33m | 527 | Not reported | Closed |
| 22 | Claude Opus 5.5low effort | 2.4% | $16.95 | 25m | 496 | Not reported | Closed |
| 23 | Grok 4.7medium effort | 2.1% | $15.91 | 32m | 437 | Not reported | Closed |
| 24 | Muse Spark 1.3medium effort | 1.5% | $5.46 | 32m | 235 | Not reported | Closed |
| 25 | Gemini 3.8 Flashmedium effort | 1.0% | $2.01 | 17m | 219 | Not reported | Closed |
| 26 | Muse Spark 1.3low effort | 0.9% | $1.89 | 16m | 98 | Not reported | Closed |
| 27 | Gemini 3.8 Flashhigh effort | 0.5% | $2.72 | 23m | 300 | Not reported | Closed |
| 28 | Gemini 3.8 Flashlow effort | 0.2% | $1.66 | 11m | 227 | Not reported | Closed |
| 29 | Muse Spark 1.3high effort | 0.2% | $4.20 | 26m | 179 | Not reported | Closed |
Each model uses its vendor’s coding harness. Costs are USD per run. Full score and steps retain the source labels; the announcement does not give an exact score formula or a repeat count.
Full scores for each of the nine apps
| Model / effort | Gmail | Figma | Sheets | Slack | DocuSign | Jira | Shopify | Craigslist | Zendesk | Parameters (B) | Open / closed |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5xhigh effort | 65.0% | 14.0% | 32.5% | 36.7% | 51.2% | 8.6% | 41.8% | 18.4% | 39.7% | Not reported | Closed |
| Claude Opus 5.5max effort | 65.8% | 14.2% | 18.4% | 34.0% | 48.9% | 20.4% | 13.1% | 24.5% | 51.8% | Not reported | Closed |
| Claude Opus 5.5high effort | 71.8% | 7.6% | 27.4% | 20.3% | 41.5% | 9.3% | 21.9% | 20.8% | 53.5% | Not reported | Closed |
| Claude Opus 5.5medium effort | 42.3% | 13.3% | 21.9% | 13.0% | 22.1% | 5.2% | 13.6% | 18.6% | 17.4% | Not reported | Closed |
| GPT-6 Astramax effort | 33.8% | 4.6% | 8.2% | 13.9% | 8.7% | 9.2% | 38.7% | 7.8% | 32.6% | Not reported | Closed |
| Claude Fable 5.1high effort | 57.6% | 2.2% | 2.6% | 7.6% | 9.2% | 6.3% | 11.5% | 14.7% | 7.7% | Not reported | Closed |
| Claude Fable 5.1xhigh effort | 28.2% | 2.6% | 6.3% | 7.6% | 7.1% | 6.8% | 22.1% | 20.9% | 15.3% | Not reported | Closed |
| GPT-6.1 Solmax effort | 30.4% | 3.3% | 7.1% | 12.2% | 13.2% | 5.4% | 16.6% | 9.8% | 15.1% | Not reported | Closed |
| GPT-6 Astraxhigh effort | 36.1% | 1.9% | 4.7% | 11.7% | 7.9% | 6.8% | 16.3% | 10.4% | 10.4% | Not reported | Closed |
| Claude Fable 5.1max effort | 34.7% | 2.1% | 4.2% | 7.3% | 5.8% | 4.0% | 14.1% | 19.6% | 12.3% | Not reported | Closed |
| GPT-6 Astrahigh effort | 11.0% | 2.7% | 6.7% | 10.1% | 14.4% | 4.7% | 13.0% | 10.7% | 16.3% | Not reported | Closed |
| GPT-6.1 Solhigh effort | 25.9% | 0.7% | 4.4% | 13.6% | 6.2% | 6.3% | 12.0% | 11.0% | 9.1% | Not reported | Closed |
| GPT-6 Astramedium effort | 17.3% | 3.8% | 6.3% | 12.0% | 6.6% | 5.3% | 11.3% | 8.8% | 15.1% | Not reported | Closed |
| Claude Fable 5.1low effort | 28.9% | 1.5% | 4.9% | 11.7% | 4.5% | 2.0% | 5.7% | 15.6% | 8.4% | Not reported | Closed |
| Claude Fable 5.1medium effort | 18.0% | 2.0% | 5.4% | 7.2% | 6.6% | 5.2% | 13.1% | 20.9% | 3.7% | Not reported | Closed |
| GPT-6.1 Solxhigh effort | 15.6% | 6.2% | 5.1% | 7.2% | 3.3% | 6.9% | 9.0% | 11.3% | 12.6% | Not reported | Closed |
| GPT-6 Astralow effort | 23.9% | 1.2% | 4.9% | 8.2% | 4.1% | 4.0% | 7.8% | 2.1% | 14.0% | Not reported | Closed |
| GPT-6.1 Solmedium effort | 14.4% | 3.7% | 3.0% | 13.2% | 3.3% | 4.5% | 8.9% | 3.0% | 8.0% | Not reported | Closed |
| GPT-6.1 Sollow effort | 5.1% | 1.5% | 3.0% | 8.8% | 3.3% | 3.0% | 1.6% | 7.5% | 7.2% | Not reported | Closed |
| Grok 4.7low effort | 12.4% | 0.0% | 0.7% | 5.3% | 0.4% | 0.9% | 2.3% | 0.8% | 2.6% | Not reported | Closed |
| Grok 4.7high effort | 11.8% | 0.2% | 1.9% | 4.6% | 0.2% | 1.4% | 1.9% | 0.8% | 1.4% | Not reported | Closed |
| Claude Opus 5.5low effort | 7.7% | 0.0% | 1.7% | 5.8% | 0.2% | 1.3% | 0.0% | 0.0% | 4.5% | Not reported | Closed |
| Grok 4.7medium effort | 9.6% | 0.4% | 0.0% | 4.8% | 0.2% | 0.0% | 3.0% | 0.0% | 1.1% | Not reported | Closed |
| Muse Spark 1.3medium effort | 8.7% | 0.0% | 0.0% | 4.9% | 0.4% | 0.0% | 0.0% | 0.0% | 0.0% | Not reported | Closed |
| Gemini 3.8 Flashmedium effort | 2.4% | 0.0% | 0.4% | 5.2% | 0.4% | 0.0% | 1.1% | 0.0% | 0.0% | Not reported | Closed |
| Muse Spark 1.3low effort | 0.9% | 0.0% | 0.4% | 4.9% | 0.2% | 0.0% | 0.0% | 0.0% | 1.7% | Not reported | Closed |
| Gemini 3.8 Flashhigh effort | 2.4% | 0.2% | 0.4% | 0.0% | 0.4% | 0.0% | 0.0% | 0.0% | 1.1% | Not reported | Closed |
| Gemini 3.8 Flashlow effort | 1.1% | 0.0% | 0.0% | 1.1% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | Not reported | Closed |
| Muse Spark 1.3high effort | 1.8% | 0.0% | 0.0% | 0.0% | 0.2% | 0.0% | 0.0% | 0.0% | 0.0% | Not reported | Closed |
How to read this leaderboard
Compare effort settings alongside their run cost and time. The published best setting is not always the highest effort: Claude Opus 5.5 scores 34.2% at xhigh and 32.3% at max. Costs describe complete runs in this benchmark, so they are separate from API token pricing.
Operator receipt: 29 sourced rows are currently displayable on this page; the leading published row is Claude Opus 5.5 (xhigh) at 34.2%.
Honest limit: These results combine a model, its vendor's harness, and a reasoning-effort setting. The announcement does not disclose an exact full-score formula, evaluated API snapshot IDs, run dates, repeat counts, or uncertainty intervals. We keep this benchmark outside weighted model rankings.
How we show Web Clone Bench
We mirror Plato’s Web Clone Bench v1 tables: 29 effort configurations across 7 models and 9 web apps. Agents explore a running reference and rebuild it from an empty directory. End-to-end workflows check both visual and functional behavior.
Each row includes the vendor’s coding harness and reasoning-effort choice. The table keeps the published full score, average cost and time per run, steps, and per-app scores together. These results remain outside weighted model rankings.
Snapshot
The published Web Clone Bench snapshot places Claude Opus 5.5 (xhigh) first at 34.2%. The third row is 3.7 points behind. The broader top-10 range is 22.6 points, so the table still separates the published systems.
29 effort configurations are shown for Web Clone Bench. The benchmark falls in the Coding category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. Web Clone Bench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Web Clone Bench
Year
2026
Tasks
9 reference web apps
Format
Published full score (%)
Difficulty
Long-running web application recreation
Each agent explores a running reference app and builds its own replica without starting code. Runs use each vendor's coding harness and end at submission or an eight-hour limit. We preserve all 29 published model-and-effort configurations across seven models, including average run cost, time, steps, and scores for each app.
Recreating a page is one frontend task. The app-development model shortlist keeps that evidence alongside other checks for the job you need.
Freshness and provenance
Version
Web Clone Bench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Web Clone Bench measure?
Coding agents recreate nine reference web apps from an empty directory. Plato checks visual and functional behavior through end-to-end workflows.
Which model leads the published Web Clone Bench snapshot?
Claude Opus 5.5 (xhigh) currently leads the published Web Clone Bench snapshot with 34.2% full score (%). BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on Web Clone Bench?
The October 15, 2026 snapshot snapshot contains 29 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.