ViBench 2.0: A Benchmark for Long-Horizon Software Building
Comments on “ViBench 2.0: A Benchmark for Long-Horizon Software Building”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Replit released ViBench 2.0, a benchmark in which an agent starts from an empty project and builds one of seventeen products, adding up to eighteen features in sequence (241 requests, 544 browser-graded test plans). Per Replit, the top three models pass every plan in only 49–65% of builds, and failures cluster where features interact. Results come from Fireworks' runs in one OpenHands-based harness, so other setups may score differently. Several figures, including cost, speed and the grader check, are Replit's own and were not independently verified.
Why it ranked
No. 9 of 59 with a score of 0.66. The panel rated it highest on new information (0.91) and substantiated (0.87), lowest on practical relevance (0.65). Placed No. 9, No. 13 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Benchmark release for AI coding agents with substantial methodology and results, partly vendor-run and promotional in tone, with a tested grader and public repos supporting claims.
- AI judgeJudge 3gpt-5.6-luna
- A new, data-rich benchmark release with public artifacts, test methodology, reproducible runs, model comparisons, costs, and failure analysis. It is highly relevant to AI coding agents and software builders, with modest promotional framing but limited overclaiming.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.66total under rubric v2.3
- Significance0.67
- Reader interest0.78
- Audience breadth0.69
- Practical relevance0.65
- Substantiated0.87
- Technical depth0.76
- New information0.91
- Counts against: Promotional0.22
- Counts against: Overclaim0.19
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.68 | 0.55 | 0.78 | 0.67 | 0.23 | +0.24 | +0.161 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.71 | 0.70 | 0.93 | 0.78 | 0.23 | +0.26 | +0.203 |
| Audience breadthHow many people who build technology does it concern? | 0.74 | 0.60 | 0.72 | 0.69 | 0.14 | +0.13 | +0.089 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.62 | 0.55 | 0.78 | 0.65 | 0.23 | +0.22 | +0.143 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.98 | 0.72 | 0.90 | 0.87 | 0.26 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 0.81 | 0.68 | 0.80 | 0.76 | 0.13 | +0.05 | +0.038 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.85 | 0.99 | 0.91 | 0.14 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.11 | 0.35 | 0.20 | 0.22 | 0.24 | −0.19 | −0.042 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.21 | 0.25 | 0.12 | 0.19 | 0.13 | −0.11 | −0.021 |
| Sum of contributions | +0.660 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.660 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-01/9/"><img src="https://rubric.hrih.si/badge/2026-11-01/9.svg" alt="Ranked #9 on Rubric, 1 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 2 matches source · 4 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
ViBench 2.0 consists of seventeen products, 241 feature requests, and 544 test plans.
The cited source says the same thing.
The article says “ViBench 2.0 has seventeen products, 241 feature requests, and 544 test plans”
The source says “each of 17 real-world apps one feature at a time. A grading agent then tests each app in a browser, the way a user would. In total, ViBench 2.0 has…”
Source cited by the article: github.com/ViBench/vibench-public
Matches sourceBenchmark result
The best models, GPT-6.1 Sol, GPT-6 Astra, and Opus 5.5, pass all test plans in only 49-65% of their builds.
The cited source says the same thing.
The article says “The best models, GPT-6.1 Sol, GPT-6 Astra, and Opus 5.5, pass all test plans in only 49-65% of their builds”
The source says “3 | Claude Opus 5.5 | Anthropic | 48.5% | $23.05 | 4,380”
Source cited by the article: fireworks.ai/specialized-intelligence-index/benchmarks/application-development/vibench
No source givenNumber
When GPT-6.1 Sol re-graded 477 of Opus's passes, 99.2% were confirmed correct.
The article gives no link or citation for it.
The article says “When GPT-6.1 Sol re-graded 477 of Opus’s passes, 99.2% were confirmed correct”
The source says No source to compare with.
Couldn’t checkBenchmark result
GPT-6.1 Sol leads with a pass@1 of 65%, followed by GPT-6 Astra at 60% and Opus 5.5 at 49%.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “GPT-6.1 Sol leads with a pass@1 of 65%, followed by GPT-6 Astra at 60% and Opus 5.5 at 49%”
The source says No passage could be compared.
Source cited by the article: fireworks.ai/specialized-intelligence-index/benchmarks/application-development/vibench
Couldn’t checkBenchmark result
DeepSeek V4.1 Flash passes 77% of its test plans for about $2.50 per build, though no open-weight model passes a full build.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “DeepSeek V4.1 Flash passes 77% of its test plans for about $2.50 per build, though no open-weight model passes a full build”
The source says No passage could be compared.
Source cited by the article: fireworks.ai/specialized-intelligence-index/benchmarks/application-development/vibench
No source givenCausal conclusion
Opus 5.5 finishes a build in about 70 minutes on average, roughly 2.5 times faster than Sol, and the gap mostly comes from API provider speed.
The article gives no link or citation for it.
The article says “Opus 5.5 finishes a build in about 70 minutes on average, roughly 2.5 times faster than Sol. Opus produces only about 6% fewer tokens, so the gap mostly comes from…”
The source says No source to compare with.
No source givenBenchmark result
Pass@1 rises across GPT generations from 0% for GPT-5.4 to 31% for GPT-6 Sol and 53% for GPT-6.1 Sol, while plan pass rate climbs from 42% to 95%.
The article gives no link or citation for it.
The article says “Pass@1 rises across generations, from 0% for GPT-5.4 to 31% for GPT-6 Sol and 53% for GPT-6.1 Sol, and plan pass rate climbs from 42% to 95%”
The source says No source to compare with.
No source givenNumber
Nearly three-quarters of the top models' failed builds fail only on feature intersections.
The article gives no link or citation for it.
The article says “Nearly three-quarters of those models’ failed builds fail only on these intersections”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 1 Nov 2026, 05:47 UTC. It played no part in the ranking.