Changelog

A record of feature, data, and experience improvements released to Leaderboard.cn users.

v0.1.522026-09-05

GPT-6 Astra and Complete Official Evaluations

Added GPT-6 Astra model details, API pricing, and the complete evaluation table from OpenAI's official release page.

Added GPT-6 Astra with a 1.05M input context, 128K maximum output, text-and-image input, tool calling, and reasoning, plus four official API price tiers covering Standard and Fast processing across short and long contexts.
Transcribed the complete OpenAI release table as 40 metrics and 149 available scores, preserving task versions, tool configurations, the reasoning-effort selection rule, and supplemental lower-cost, pass@4, and timing results.
Classified all five AutomationBench scores under the official leaderboard's private held-out set, retaining Astra's reported 41.4% and the unspecified-version caveat so they are not compared with the 600 public tasks.

v0.1.512026-09-03

AutomationBench Private Evaluation Split

Corrected the evaluation scope of the AutomationBench results on the Claude Fable 5.1 release page so they are not compared with the public task set.

Anthropic's system card confirms that these results use the official AutomationBench leaderboard's private held-out set; moved the snapshot's four scores from the general AutomationBench metric into a separate AutomationBench (Private) metric.
Preserved all four model scores and added provenance for the strict task pass rate, Fable 5 tasks scored as zero when safeguards intervened, and the incompatibility with the 600 public tasks.

v0.1.502026-09-03

GLM-5.3 Open Weights and New Evaluations

Synchronized GLM-5.3's open weights and added the latest Qwen3.8-Max-0902 and Muse Spark 1.3 evaluations.

Marked GLM-5.3 as open-weight and added its custom GLM-5.3 License plus the official Hugging Face and ModelScope repositories; the default repository is FP8, with BF16 weights published separately.
Added official vLLM 0.28.0+ and SGLang deployment recipes for 8×H200: the vLLM recipe includes MTP, tool-call, and reasoning parsers, while SGLang uses its verified low-latency EAGLE configuration.
Added all 17 new results from the Qwen3.8-Max-0902 performance chart to the existing Qwen3.8-Max entry without creating a new model; retained the unversioned baseline column's 16 available values as raw comparisons and stored 63 scores across the other columns.
Added Muse Spark 1.3 with its 1M context, image/video/document inputs, and Meta Model API pricing, plus all 11 metrics and 42 available scores from the official performance chart.

v0.1.492026-09-03

Arena.ai Leaderboard Refresh

Refreshed all twelve Arena.ai boards and reviewed newly prominent model identities.

Updated all twelve Arena.ai boards with 1,361 scores, using the latest source cutoff snapshots for Agent, Text, WebDev, and the other boards.
Explicitly mapped Muse Spark 1.2 xHigh, DeepSeek-V4-Pro-0813 High, and Qwen3.8-Max-0902; kept Hy4 preview and Inkling-Small as reviewed source-only identities to prevent incorrect merges.

v0.1.482026-09-03

Gemini 3.8 Flash and Official Evaluations

Added Gemini 3.8 Flash with its complete official performance table.

Added Gemini 3.8 Flash with a 1M input context, 64K output, five input modalities, three reasoning levels, and all four introductory API price tiers through December 31, 2026; marked it as Gemini 3.7 Flash's successor.
Transcribed DeepMind's complete six-model, 13-row source table as 15 metrics and 85 available scores after splitting the compound LVBench and BioMysteryBench results, with per-metric methodology retained.

v0.1.472026-09-02

Claude Fable 5.1 and Benchmark Description Updates

Added Claude Fable 5.1 with its complete official evaluation and removed benchmark descriptions that become stale as leaderboards change.

Added the Claude Fable 5.1 model; Mythos 5.1 is consolidated as the same model per Anthropic's description instead of receiving a separate entry.
Transcribed Anthropic's complete main comparison table as nine metrics and 32 available scores, preserving compound evaluation settings and raw merged-mode details.
Removed model-specific scores and time-sensitive claims such as current-best, SOTA, and most-comprehensive from 24 legacy benchmark descriptions while retaining benchmark purpose, methodology, and fixed human baselines.

v0.1.462026-08-27

Qwen3.8-Flash-Next, GLM-5.3-Flash, and DeepSeek Vision Experimental

Added three models and their complete official evaluations.

Added Qwen3.8-Flash-Next with a 125B main model plus 51B N-gram embeddings, 6B activated parameters per token, native 262K / YaRN 1M context, and a GDN + QSA hybrid architecture.
Transcribed Qwen's complete text and multimodal comparison tables as 27 metrics and 109 scores; the base-model comparison table is excluded.
Added GLM-5.3-Flash with MIT open weights, 320B-A18B parameters, a 1M context, 128K maximum output, four input modalities, and limited-time official API pricing.
Added DeepSeek-V4-Flash-Vision-Exp with a 1M context, 384K maximum output, text-and-image input, thinking/non-thinking modes, and official peak/off-peak pricing.
Transcribed the GLM main-model table as 14 metrics and 70 scores, plus the DeepSeek official chart as 11 metrics and 31 scores; the GLM base-model table is excluded.

v0.1.452026-08-25

DeepSeek-V4 and GPT-5.6 Sol Pricing

Updated DeepSeek-V4 time-based rates and GPT-5.6 Sol promotional and Fast pricing.

Updated DeepSeek-V4-Pro-0813 and DeepSeek-V4-Flash-0731 to the official peak and off-peak rate tiers, with matching English time labels.
Applied GPT-5.6 Sol promotional pricing available at least through November 21, 2026, including Standard and Fast tiers for ≤272K and >272K requests.

v0.1.442026-08-17

Meta Muse Spark 1.2 and Official Evaluations

Added the Meta Muse Spark 1.2 model and its five official evaluation charts.

Added the Muse Spark 1.2 entry: 1M context, four input modalities, and Meta Model API pricing ($1.25/$4.25, cached hits $0.15); marked 1.1's successor.
Transcribed the five official evaluation charts (Terminal-Bench 2.1, DeepSWE 1.1, GDPVal-AA v2, MCP Atlas, Meta Internal Coding Bench) across ten models, 30 scores in total.
Added the Muse Spark 1.2 DeepMind performance-table row (six metrics) to the Gemini 3.7 Flash snapshot, raising that snapshot from 78 to 84 scores.

v0.1.432026-08-17

Gemini 3.7 Flash and Official Evaluations

Added the Gemini 3.7 Flash model and its complete official evaluations.

Added the 1M context, five input modalities, and introductory-pricing notes; marked 3.6 Flash's successor.
Transcribed the complete 20-metric comparison table as 78 scores.

Home