OpenGPTAI decision workspaceLMArena
The complete latest style-controlled overall ranking, based on users comparing two model answers side by side.
- Leaderboard publication date
- Sep 2, 2026
- Data retrieved
- Sep 9, 2026
- License
- CC BY 4.0
Independent model decisions
AI model rankings
Shortlist exact model versions from public evidence, compare them on one basis, then verify the choice on your own task.
- Rank 1Claude Fable 5.1 Max EffortAnthropic83.41
- Rank 2Claude Fable 5 Max EffortAnthropic82.97
- Rank 3GPT-6 Astra Max EffortOpenAI82.16
Published evaluation data2 external public sourcesLiveBench · Edition data cutoff
OpenGPT organizes published third-party evaluation data; it did not rerun these tests. Use the rankings to shortlist models, then validate them on your own work.
Choose a ranking metric
Start with the decision you need to make. Each scenario opens the most relevant published ranking signal. The original source and scale remain visible.
Show every publisher record with its source model ID or source name, including aliases and run configurations. Source ranks are preserved; ties can still cause rank numbers to skip.
Objective-task score. LiveBench overall score across 23 objectively graded tasks in seven categories.
Showing 10 of 10. Objective-task score.
- Rank 1
Claude Fable 5.1 Max Effort
AnthropicObjective-task score83.41View source
- Rank 2
Claude Fable 5 Max Effort
AnthropicObjective-task score82.97View source
- Rank 3
GPT-6 Astra Max Effort
OpenAIObjective-task score82.16View source
- Rank 4Objective-task score81.59
View source
- Rank 5
GPT-5.6 Sol Max Effort
OpenAIObjective-task score81.05View source
- Rank 6Objective-task score80.19
View source
- Rank 7
Claude 5 Opus Thinking Max Effort
AnthropicObjective-task score80.08View source
- Rank 8
Kimi K3
Moonshot AIObjective-task score79.19View source
- Rank 9
Gemini 3.7 Flash High
GoogleObjective-task score78.83View source
- Rank 10
Qwen 3.8 Max
AlibabaObjective-task score78.46View source
Independent sourceLiveBench10 source-listed model entries · Data retrieved: Sep 9, 2026
Ranking methodEach view preserves the source’s metric, release, and methodology. Follow the links to audit the underlying work.
Ranking basis
Full published ranking and official model IDs
The main ranking preserves every publisher record and shows its source model ID or source name. Every row can be compared: exact, configuration-safe mappings use official versions; all other rows stay within one source ranking. Missing data never becomes a zero score.- 79 official model IDs
- 79
- 0 with public ranking data
- 0
- 79 awaiting comparable data
- 79
- LiveBench · 56 source records retained for audit
- 56
Ranking basis
How to read this research edition
Scores stay on their original publisher scales. They are not combined into a composite score.
Missing data does not mean zero. It means the publisher did not report that metric for this snapshot.
When a source reports model settings, the position applies to that exact setup and source release—not every task, region, or deployment.
No provider can pay for placement. OpenGPT links directly to the evidence used.
Ranking basis
Data and sources
Each view preserves the source’s metric, release, and methodology. Follow the links to audit the underlying work.LiveBench
Twenty-three objective tasks across seven categories; questions refresh every six months.
- Source snapshot
- LiveBench-2026-06-25
- Data retrieved
- Sep 9, 2026
- License
- Apache-2.0 project license; exact new-livebench publication repository does not repeat a LICENSE file
Method version
A predictable publishing rhythm
OpenGPT checks active model sources every six hours. LMArena and LiveBench publish only after their versioned validation policies pass. Paused sources remain historical and do not enter current rankings.- 01
6-hour source checks
LMArena and LiveBench updates publish automatically only after schema, completeness, date, integrity, and anomaly checks. Paused sources stay historical.
- 02
Weekly change summary
Summarize meaningful ranking, model, and source changes in one reviewable update.
- 03
Monthly history snapshot
Freeze a dated snapshot with exact configurations, source releases, and corrections.