AI Battle > Rankings > GDPval-AA v2: Average Turns per Task
GLM-5.3-Flash by Z AI leads the GDPval-AA v2: Average Turns per Task ranking with a score of 80.9. The ranking covers 20 evaluated eval gdpval from a public dataset that AI Battle re-checks daily.
Last updated: . Source: public benchmark datasets and provider pricing, aggregated daily; 20 entries shown.
| # | Entry | Creator | Value | Δ vs #1 |
|---|---|---|---|---|
| 1 | GLM-5.3-Flash | Z AI | 80.9 | 0.0% |
| 2 | Gemini 3.8 Flash (high) | 78.2 | -3.3% | |
| 3 | Muse Spark 1.3 (max) | Meta | 74.5 | -8.0% |
| 4 | DeepSeek V4.1 Flash (Reasoning, Max Effort) | DeepSeek | 72.1 | -10.9% |
| 5 | GLM-5.3 (max) | Z AI | 70.0 | -13.5% |
| 6 | Qwen3.8 2.4T A95B | Alibaba | 67.0 | -17.3% |
| 7 | Muse Spark 1.3 (xhigh) | Meta | 63.1 | -22.0% |
| 8 | Claude Opus 5 (Adaptive Reasoning, Max Effort) | Anthropic | 60.5 | -25.3% |
| 9 | Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) | Anthropic | 60.4 | -25.4% |
| 10 | Qwen3.8 27B (xhigh) | Alibaba | 58.0 | -28.3% |
| 11 | MiniMax-M3 | MiniMax | 54.2 | -33.1% |
| 12 | Claude Fable 5.1 (Adaptive Reasoning, Xhigh Effort, Default Fallback) | Anthropic | 50.7 | -37.3% |
| 13 | Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) | Anthropic | 50.7 | -37.4% |
| 14 | GPT-5.6 Sol (max) | OpenAI | 45.2 | -44.1% |
| 15 | K2 Horizon 375B A23B | MBZUAI Institute of Foundation Models | 45.0 | -44.4% |
| 16 | Inkling (xhigh) | Thinking Machines | 43.1 | -46.7% |
| 17 | Grok 4.6 (xhigh) | SpaceXAI | 41.9 | -48.2% |
| 18 | DeepSeek V4 Pro 0813 (Reasoning, Max Effort) | DeepSeek | 38.9 | -51.9% |
| 19 | Kimi K3 (max) | Kimi | 37.5 | -53.6% |
| 20 | Grok 4.6 (high) | SpaceXAI | 37.0 | -54.3% |
Across the 20 numeric entries in this window the median is 56.1 and the mean is 56.4, so the typical entry sits 24.9 below the leader. The window spans 37.0 to 80.9.
10 of 20 entries are at or above the median, and the top three hold 21% of the combined value — read it against the metric below, since a higher number is not "better" for every dataset.
12 creators appear in this window; the three most-represented hold 40% of the entries.
| Creator | Entries | Highest value | Top entry |
|---|---|---|---|
| Anthropic | 4 | 60.5 | Claude Opus 5 (Adaptive Reasoning, Max Effort) |
| Z AI | 2 | 80.9 | GLM-5.3-Flash |
| Meta | 2 | 74.5 | Muse Spark 1.3 (max) |
| DeepSeek | 2 | 72.1 | DeepSeek V4.1 Flash (Reasoning, Max Effort) |
| Alibaba | 2 | 67.0 | Qwen3.8 2.4T A95B |
| SpaceXAI | 2 | 41.9 | Grok 4.6 (xhigh) |
| 1 | 78.2 | Gemini 3.8 Flash (high) | |
| MiniMax | 1 | 54.2 | MiniMax-M3 |
| OpenAI | 1 | 45.2 | GPT-5.6 Sol (max) |
| MBZUAI Institute of Foundation Models | 1 | 45.0 | K2 Horizon 375B A23B |
| Thinking Machines | 1 | 43.1 | Inkling (xhigh) |
| Kimi | 1 | 37.5 | Kimi K3 (max) |
The same models are measured across multiple independent datasets. Aggregating the 20 entries above against every other AI Battle dataset gives this cross-section, computed by AI Battle and published nowhere upstream:
| Model | Datasets | Best placements elsewhere |
|---|---|---|
| GLM-5.3-Flash | 54 | #1/20 in Pricing: Cache Hit, Input, and Output, #1/20 in Pricing: Cache Hit, Input, and Output, #1/20 in AA-Omniscience Index: Cost Breakdown |
| Gemini 3.8 Flash (high) | 60 | #1/11 in Speed, #1/11 in Speed, #2/20 in Output Speed |
| Muse Spark 1.3 (max) | 61 | #3/20 in AA-Briefcase Elo, #3/20 in Output Speed, #3/20 in GDPval-AA v2 Leaderboard |
| DeepSeek V4.1 Flash (Reasoning, Max Effort) | 56 | #3/20 in End-to-End Response Time, #3/20 in Latency: Time To First Answer Token, #2/11 in Cost per Task |
| GLM-5.3 (max) | 57 | #3/20 in AA-Briefcase Tool Calls Breakdown, Avg per Task, #4/20 in GDPval-AA v2 Leaderboard, #4/20 in AA-Omniscience Hallucination Rate |
| Qwen3.8 2.4T A95B | 43 | #2/14 in Model Size: Total and Active Parameters, #5/20 in AA-Briefcase Tool Calls Breakdown, Avg per Task, #5/20 in 𝜏³-Banking: Score |
| Muse Spark 1.3 (xhigh) | 12 | #3/20 in 𝜏³-Banking: Time per Task, #5/20 in SciCode: Score, #7/20 in GDPval-AA v2 Leaderboard |
| Claude Opus 5 (Adaptive Reasoning, Max Effort) | 44 | #2/20 in AA-Briefcase Elo, #2/20 in GDPval-AA v2 Leaderboard, #2/20 in AA-Briefcase Elo |
| Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) | 36 | #1/20 in AA-Briefcase Elo, #1/20 in AA-Omniscience Index, #1/20 in Artificial Analysis Intelligence Index by Open Weights / Proprietary |
| Qwen3.8 27B (xhigh) | 43 | #6/20 in AA-Omniscience Hallucination Rate, #6/20 in 𝜏³-Banking: Score, #8/20 in AA-LCR v1.1: Cost per Task |
| MiniMax-M3 | 57 | #1/20 in AA-Omniscience Hallucination Rate, #1/20 in 𝜏³-Banking: Output Tokens per Task, #4/20 in Pricing: Cache Hit, Input, and Output |
| Claude Fable 5.1 (Adaptive Reasoning, Xhigh Effort, Default Fallback) | 20 | #2/20 in Humanity's Last Exam: Score, #2/20 in AA-Omniscience Accuracy, #2/20 in AA-Briefcase Rubric Pass Rate by File Type (Normalized) |
| Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) | 11 | #4/20 in AA-Briefcase Rubric Pass Rate by File Type (Normalized), #4/20 in AA-Briefcase Elo, #4/20 in GDPval-AA v2 Leaderboard |
| GPT-5.6 Sol (max) | 56 | #3/20 in SciCode: Output Tokens per Task, #4/20 in 𝜏³-Banking: Output Tokens per Task, #5/20 in AA-LCR v1.1: Output Tokens per Task |
| K2 Horizon 375B A23B | 24 | #2/20 in AA-Omniscience Hallucination Rate, #8/20 in AA-Briefcase Tool Calls Breakdown, Avg per Task, #12/20 in AA-Omniscience Index: Token Usage |
| Inkling (xhigh) | 52 | #3/20 in Context Window, #3/20 in 𝜏³-Banking: Output Tokens per Task, #5/20 in AA-Briefcase Output Tokens per Task |
| Grok 4.6 (xhigh) | 3 | #6/20 in GDPval-AA v2 Leaderboard, #15/20 in GDPval-AA v2: Output Tokens per Task |
| DeepSeek V4 Pro 0813 (Reasoning, Max Effort) | 59 | #3/14 in Model Size: Total and Active Parameters, #3/12 in Artificial Analysis Openness Index: Score, #3/11 in Cost per Task |
| Kimi K3 (max) | 60 | #1/20 in Context Window, #1/20 in AA-LCR v1.1: Score, #1/14 in Model Size: Total and Active Parameters |
| Grok 4.6 (high) | 61 | #2/20 in 𝜏³-Banking: Score, #5/20 in AA-Briefcase Elo, #5/20 in AA-Omniscience Index |
This ranking carries the metric
averageTurnsPerTask from the
GDPval-AA v2: Average Turns per Task dataset in the tab::eval-gdpval family. Units and the
direction of "better" are not declared by the source dataset, so AI Battle does not invent
them: values are shown exactly as published, and the ordering is the source dataset's own.
Rebuilt on from the
public dataset GDPval-AA v2: Average Turns per Task (family tab::eval-gdpval).
The source lists 20 entries and all 20 are shown here. AI Battle requests
the dataset's rows, keeps the ordering the source publishes, and computes only the statistics
on this page: median, mean, spread, the #1–#2 gap, the creator breakdown and the raw-data file.
Nothing on this page is generated text about the entries — the numbers come from the source.
Refresh cadence: every deploy re-pulls the dataset, so this page tracks the source's own update cycle. Corrections: see the methodology page.
Machine-readable copy of this table, with the statistics below and the source provenance: /data/rankings/eval-gdpval-gdpval-aa-v2-average-turns-per-task.json. Free to reuse with attribution to AI Battle and the source dataset.
GLM-5.3-Flash by Z AI currently leads the GDPval-AA v2: Average Turns per Task ranking with a score of 80.9, the highest value in this dataset across eval gdpval. AI Battle aggregates the dataset from its public source and re-checks it daily.
The GDPval-AA v2: Average Turns per Task ranking is rebuilt daily from the upstream public dataset. Scores are the source dataset's own published values, so positions track whatever that source last released — each ranking page names the dataset it comes from.
The top 3 are: 1. GLM-5.3-Flash (80.9), 2. Gemini 3.8 Flash (high) (78.2), 3. Muse Spark 1.3 (max) (74.5).