Skip to content

Zhipu AI · Chat

GLM 5.3

Zhipu AI 🇨🇳ChatGLM familyOpen weights. One of its sources says the weights are publicly available.
From Cloudflare Workers AI's listing:
GLM-5.3 is Z.ai's flagship agentic coding model, pairing a 1M-token context window with reasoning, function calling, and structured outputs to power multi-step, tool-driven development workflows.
llminfo IQ
138.2
Market median price
$2.15 blended / 1M
Max context
1.3M
Listings
361

Evidence and sources

🤗 zai-org/GLM-5.3 Model architecture, as its sources state it llminfo IQ 138.2, ranked among chat models. From 318 benchmarks. Plenty of data behind it.How llminfo IQ is computedCoding IQ 121.0, 95% interval 112–130. From 26 benchmark families across 8 sources. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.How Coding IQ is computedGood value (57.7). Cheaper per task than its benchmark results predict. From 51 benchmarks with measured task costs. Plenty of data behind it.How Value for money worksYap score: tokens used per coding task vs. models with the same resultsChatty (65 ±10). Generates more tokens than models with similar coding results. Measured at max effort. A fair amount of data behind it. By effort: default 56 · max 65.Median 112,175 output tokens per task at max effort.How the Yap score worksReleased 14 Aug 2026Maker HQ: CN (source)
All rankings: #5 on NanoGPT this week #6 on OpenRouter this month #12 on OpenRouter today #13 on OpenRouter this week #19 on NanoGPT recentlySources: Arcee AI data, Cloudflare Workers AI, EmpirioLabs data, Models.dev data, llminfo researchPrimary data from Arcee AI from Cloudflare Workers AI: architecture, description from EmpirioLabs: primary mode from Models.dev: weights availability from llminfo research: release date
Tags from its sources:instructreasoningtoolsopen-sourcelong-contextcodingagenticfunction-calling

Advantages and trade-offs

Compared with 10 chat models, picked mainly by popularity and IQ

Advantages

We only make a claim when there's enough data on models of the same type. Updated 22 Nov, 06:07 UTC.

Input & output

as stated by the model's providers
Text
Images
Text

Capabilities

Supported operations and provider-stated features

VisionReasoningFunction callingStructured outputPrompt cachingWeb searchBatch API

Listings

 Aihubmixcoding-glm-5.3-free · coding, freeSource: AIHubMix data docs1M$0.00$0.00
 Tokenrouterz-ai/glm-5.3-free · freeSource: Models.dev data1M$0.00$0.00
 Nous Portalz-ai/glm-5.3:batch · batchSource: Nous Portal Inference Gateway data1M$0.28cache read $0.061$1.249
 OpenRouterz-ai/glm-5.3:batch · batchSource: OpenRouter data docs1M$0.28cache read $0.061$1.249
 Kenariglm-5-3Source: Models.dev data1M$0.2857cache read $0.0531$3.43
 Uplinkswitchglm-5.3Source: Models.dev data1M$0.3178cache read $0.066$0.9987
 Cloudferro Sherlockzai-org/GLM-5.3Source: Models.dev data262K$0.34$4.48
 Eurouterglm-5.3Source: EURouter data1M$0.50cache read $0.15$2.00
 Hugging FaceInference by Coherezai-org/GLM-5.3Source: Hugging Face Inference Providers data1M$0.50$3.50
 Llmgatewayglm-5.3Source: LLMGateway data1M$0.50cache read $0.15$2.00
Show all 340 listings

Showing 10 of 340: the cheapest per 1M input tokens from each provider. 27 listings with suspiciously low prices are in the full list only.

Also sold by 302ai, Abliteratedllm, Adzuki, Aiand, Aimlapi, Akash, Alibaba Cloud, Alibaba Cloud (China), Amazon Web Services, Apertus, Arcee, Atlascloud, Avian, Baidu, Baseten, Bothub, Boundless, Cheapaimodel, Cheapest Token, Chutes, Cline Pass, Cloudcode, Cloudflare Workers AI, Cometapi, Commandcode, Cortecs AI, Crossmodel, Crusoe, Databricks, DeepInfra, Digitalocean, Doubleword, Eden AI, Empiriolabs, Engy, Entrim, Evolink, Ezkielyna, Fastrouter, Featherless AI, Fireworks AI, Flowdot, Friendli AI, Frogbot, Geodd, Gisoo, Gmicloud, Graphn, Greenpt, Huawei Maas, Hyper, Inceptron, Inco, Inferenceapis, Inferris, Ionet, Iteracompute, Jzstoken, Keln, Kilo, Kitani, Klavo, Korouter, Liquid, Llmgateway Providers, Lyceum, Meganova, Melious, Merge Gateway, Microsoft Azure, Mistral AI, Moark, Modal, Morph, Nan, NanoGPT, Near AI, Nebius AI Studio, Neon, Neosantara, Netmind, Neuralwatt, Novita AI, Obscurify, Ofox, Ollama, Opencode Go, Opper, Orcarouter, Perplexity, Perplexity Agent, Phala, Ppio, Privatemode AI, Proxystacks, Qiniu, Redpill, Runcrate, Runinfra, SambaNova, Sandbase, Scaleway, Scx AI, Sference, SiliconFlow, Siliconflow Cn, Standardcompute, Streamlake, Synthetic, Synthorai, T Systems, Teamorouter, Tencent, Tensorix, Tensorx, Tenstorrent, Tinfoil, Together AI, Tokengo, Tokenmix, Umans AI, Vancine, Venice, Vercel, Vivgrid, Vultr, Wafer, Wally, Watt, Wavespeed, Weights & Biases, Z.ai, Zenmux, Zhipu AI (Open Platform), Zro.

21 listings without a price are hidden.

Included in plans

CheapestInference Frontier PoolcheapestinferenceSubscription$71/moPlan page
Umans Code ProumansSubscription$20/moslows down past the allowancePlan page
Umans Code MaxumansSubscription$50/moslows down past the allowancePlan page
GLM Coding Plan LitezhipuSubscription$18/moPlan page
GLM Coding Plan ProzhipuSubscription$72/moPlan page
GLM Coding Plan MaxzhipuSubscription$160/moPlan page

Benchmarks

Named effort levels run from lowest to highest. Select a result to see its source details.

Agentic

Reported by Z.ai, DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Arena AgentSource: LMArena data
BFCL-V4Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · official BFCL V4 harness commit 6ea5797 · 5,217 cases · temperature 0.001 · weighted overall score · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepResearch Bench IIAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · Claude Code · identical search tools · four-hour task limit · GPT-5.5 judge · three independent scoring passes · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
GDPval-AA3 related benchmarks · 3 results
GDPval-AA
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
GDPval · Atria paper Elo
Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · 220 tasks · AgentCompass / OpenHands · four-hour limit · 250 steps · GLM-5.3 rubric judge · pairwise Elo
v2
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Artificial Analysis / Z.ai转载
JobBenchAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · official OpenCode harness · main split / 65 tasks · 7,200-second limit · two attempts · Grok-4.3 judge · mean task score · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Claude Code 2.1.207 · max · 3-run weighted average · documented H100 fallback
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
SWE-Bench ProAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · OpenHands v1.23.0 · 4.5-hour limit · 250 interaction steps · network disabled and Git history removed · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Source: OpenRouter benchmark exports
16 Aug 2026 version · 90-day results
Across 44 providers lowest 73.7% median 76.8% highest 78.8%OpenRouter's automatic routing 77.3%
Show results from 44 providers and OpenRouter's automatic routingHide results from 44 providers and OpenRouter's automatic routing
  • Wafer
  • Together
  • Fireworks
  • Doubleword
  • Crusoe
  • ModelRun
  • Z.AI
  • BaseTen
  • CoreWeave
  • DigitalOcean
  • AtlasCloud
  • Cloudflare
  • Friendli
  • Tenstorrent
  • DeepInfra
  • Io Net
  • StreamLake
  • OpenRouter's automatic routing
  • Nebius
  • Novita
  • Parasail
  • Venice
  • Phala
  • Decart
  • Alibaba
  • Featherless
  • Relace
  • Makora
  • Inceptron
  • Mistral
  • Modal
  • Sail Research
  • InferenceNet
  • SiliconFlow
  • PrimeIntellect
  • DekaLLM
  • Morph
  • SambaNova
  • Baidu
  • Reka
  • GMICloud
  • River
  • Cohere
  • AkashML
  • Amazon Bedrock
digitalocean · glm-5.3 · 90-day results
Terminal-Bench5 related benchmarks · 14 results
Terminal-Bench
Source: Hugging Face Official Benchmarks data · Reported by the model's developer. The source relaying it did not verify it. claim
Terminal-Bench 2.1
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Setting not stated
Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · four-run mean · 12× timeout multiplier capped at 12 hours · 500 turns · Claude Code v2.1.224 except GPT models use Terminus 2
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Source: Hugging Face Official Benchmarks data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Terminal-Bench 3.0
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Setting not stated
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek Harness / max effort
DeepSeek V4.1 launch chart · settings unspecified
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Terminal-Bench 4.0
Reported by Terminal-Bench / Stanford / Harbor / Laude Institute · via BenchBoard docs
Terminal-Bench 4.0 official leaderboard · Claude Code · max effort
Source: LLM Stats data
Setting not stated
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek Harness / max effort
DeepSeek V4.1 launch chart · settings unspecified

Multilingual

Bangla LLM BenchSource: Bangla LLM Bench data docs
Benchmark Nusantara17 related benchmarks · 17 results
Show 11 more resultsHide 11 results
EKI EstonianSource: Keelemudelite mõõdupuu data docs · Aug 2026
NepNewsBench v2 draft40 related benchmarks · 40 results
entity F1 (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
core fact recall (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
core fact recall (English-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
Show 34 more resultsHide 34 results
effective score (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
effective score (English-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
effective score (Nepali-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
English prose quality (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity F1 (English-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity F1 (Nepali-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity precision (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity precision (English-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity precision (Nepali-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity recall (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity recall (English-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
entity recall (Nepali-source questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
Nepali prose quality (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
supporting fact recall (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low
valid-answer quality (all questions)
Source: NepNewsBench v2 (draft candidate release) data · Sep 2026
Z.AI via openrouter, reasoning low

Science

GPQA Diamond49 results
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Pass@1 / max effort
Source: OpenRouter benchmark exports
16 Aug 2026 version · 90-day results
Across 45 providers lowest 81.5% median 84.3% highest 88.6%OpenRouter's automatic routing 85.4%
Show results from 45 providers and OpenRouter's automatic routingHide results from 45 providers and OpenRouter's automatic routing
  • Doubleword
  • OpenInference
  • DigitalOcean
  • Wafer
  • DeepInfra
  • Z.AI
  • InferenceNet
  • Venice
  • GMICloud
  • OpenRouter's automatic routing
  • Inceptron
  • Baidu
  • DekaLLM
  • BaseTen
  • River
  • SambaNova
  • ModelRun
  • AtlasCloud
  • Phala
  • Fireworks
  • Crusoe
  • SiliconFlow
  • PrimeIntellect
  • Relace
  • Decart
  • Mistral
  • Makora
  • Together
  • Sail Research
  • Reka
  • Nebius
  • Novita
  • Modal
  • AkashML
  • Cohere
  • Alibaba
  • Parasail
  • Amazon Bedrock
  • Cloudflare
  • Morph
  • CoreWeave
  • Friendli
  • Io Net
  • Featherless
  • StreamLake
  • Tenstorrent

Human preference

Arena ELO49 related benchmarks · 49 results
Source: LMArena data · Nov 2026
BenchmarkMax
Arena ELO
Chinese
Creative Writing
English
Show 45 more benchmarksHide 45 benchmarks
BenchmarkMax
Exclude Ties
Expert
French
German
Hard Prompts
Hard Prompts English
Industry Business and Management and Financial Operations
Industry Entertainment and Sports and Media
Industry Life and Physical and Social Science
Industry Mathematical
Industry Medicine and Healthcare
Industry Software and It Services
Industry Writing and Literature and Language
Japanese
Korean
Longer Query
Multi Turn
Non English
Polish
Russian
Spanish
Style-Controlled
Text Factuality
Text Factuality - Chinese
Text Factuality - Creative Writing
Text Factuality - English
Text Factuality - Exclude Ties
Text Factuality - Expert
Text Factuality - French
Text Factuality - Hard Prompts
Text Factuality - Hard Prompts English
Text Factuality - Industry Business and Management and Financial Operations
Text Factuality - Industry Entertainment and Sports and Media
Text Factuality - Industry Life and Physical and Social Science
Text Factuality - Industry Mathematical
Text Factuality - Industry Medicine and Healthcare
Text Factuality - Industry Software and It Services
Text Factuality - Industry Writing and Literature and Language
Text Factuality - Longer Query
Text Factuality - Multi Turn
Text Factuality - Non English
Text Factuality - Russian
Text Factuality - Spanish

Coding

Arena ELOSource: LMArena data · Nov 2026
CyberGym4 results
Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · Claude Code v2.1.119 · high thinking effort · four-hour-ten-minute limit · 2 CPUs / 4 GB
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Claude Code / 无Web工具
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
ExploitBench2 results
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
FrontierSWE2 results
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Dominance · Proximal · 1M · max effort · 128K output · snapshot 2026-08-14
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
NL2RepoSource: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
SkillsBenchLing-3.1-flash official chart · SkillsBench · version, task count, harness, reasoning setting and run count not stated · Reported by Ant inclusionAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
SWE-Marathon2 results
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
v1.1 · Claude Code 2.1.207 · max · temp 1.0 · top_p .95 · 1M · 128K output
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated

General

Long context

Reasoning

Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Setting not stated
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
AutomationBench v1.0.6Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · v1.0.6 with null-type fix from PR #13 · candidate harness and task count unstated · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified · text-only subset
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.

Math

Instruction following

Factuality

Knowledge

Other

DeepSearchQA · Atria paper scoreAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · in-house search/visit/python harness · 500 steps · 256K context / 64K max output · score subtype unstated · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSWE v1.13 results
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
mini-SWE / DeepSeek设置
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
mini-swe-agent / Z.ai设置
E-Commerce Bench7 related benchmarks · 7 results
Mean final assets
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
AnchorRatio · repeat-order pricing
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
BadSpend · fraudulent-supplier spend
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
Controllable return
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
CSE+ · honest-deal negotiation
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
Drawdown / peak assets
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
Show 1 more resultHide 1 result
Yuan per tool call
Reported by E-Commerce Bench authors / Alibaba Qwen · via BenchBoard docs
benchmark-owner 365-day merchant simulation · five independent episodes · up to four stores · initial balance ¥100,000 · max reasoning · exact API snapshot not stated for unversioned labels
ExploitGym6 related benchmarks · 13 results
Solved Tasks
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
2-hour normalized budget
6-hour normalized budget
on-target exploits
Source: ExploitGym leaderboard data docs
By eval · agent: Claude Code · version: unstated · focus: model
agent: DoGNAVY, eval: 10h timeout, selected subset, version: v1, focus: agent
agent: VARAS-OneMind (智安·宇盾), eval: 6h timeout, version: v1, focus: agent
on-target exploits (kernel)
Source: ExploitGym leaderboard data docs · Sep 2026
agent: DoGNAVY, eval: 10h timeout, selected subset, version: v1, focus: agent
agent: VARAS-OneMind (智安·宇盾), eval: 6h timeout, version: v1, focus: agent
on-target exploits (userspace)
Source: ExploitGym leaderboard data docs · Sep 2026
agent: DoGNAVY, eval: 10h timeout, selected subset, version: v1, focus: agent
agent: VARAS-OneMind (智安·宇盾), eval: 6h timeout, version: v1, focus: agent
on-target exploits (V8)
Source: ExploitGym leaderboard data docs · Sep 2026
agent: DoGNAVY, eval: 10h timeout, selected subset, version: v1, focus: agent
agent: VARAS-OneMind (智安·宇盾), eval: 6h timeout, version: v1, focus: agent
Success Rate
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
MLE-Bench Lite · HumanRankAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · Claude Code · 12-hour task limit · HumanRank score · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
NL2Repo-Bench3 results
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Setting not stated
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified · launch-page table
Reported by Z.ai · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reported by DeepSeek · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
DeepSeek V4.1 launch chart · settings unspecified
RecreationBench paper26 related benchmarks · 26 results
Web · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Web · nominal 50-task split · platform-specific score
Android · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Android · nominal 50-task split · platform-specific score
Android · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Android · nominal 50-task split · platform-specific score
Android · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Android · nominal 50-task split · platform-specific score
Android · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Android · nominal 50-task split · platform-specific score
macOS · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score
Show 20 more resultsHide 20 results
macOS · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score
macOS · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score
macOS · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score
Overall · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Overall · 250 held-out tasks · equal-weight platform macro average
Overall · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Overall · 250 held-out tasks · equal-weight platform macro average
Overall · Estimated API cost per task
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · estimated mean API cost per task · 90% cache reads assumed
Overall · Programmatic + visual mean
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Overall · 250 held-out tasks · equal-weight platform macro average
Overall · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Overall · 250 held-out tasks · equal-weight platform macro average
Overall · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Overall · 250 held-out tasks · equal-weight platform macro average
Ubuntu · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Ubuntu · nominal 50-task split · platform-specific score
Ubuntu · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Ubuntu · nominal 50-task split · platform-specific score
Ubuntu · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Ubuntu · nominal 50-task split · platform-specific score
Ubuntu · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Ubuntu · nominal 50-task split · platform-specific score
Web · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Web · nominal 50-task split · platform-specific score
Web · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Web · nominal 50-task split · platform-specific score
Web · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Web · nominal 50-task split · platform-specific score
Windows · Applications with Prog =100%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Windows · nominal 50-task split · platform-specific score
Windows · Applications with Prog ≥90%
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Windows · nominal 50-task split · platform-specific score
Windows · Programmatic assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Windows · nominal 50-task split · platform-specific score
Windows · Visual assertion score
Reported by QwenLM / RecreationWorld authors · via BenchBoard docs
RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · Windows · nominal 50-task split · platform-specific score
SkillsBench8 related benchmarks · 8 results
Text-only (79 tasks)
Reported by Tencent Hy · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Hy4 official Benchmark Appendix · highest available reasoning setting · Claude Code · 79 self-contained text-only tasks · three-run average
Text-only (79 tasks) · OpenHands
Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Atria Table 1 / Appendix B · OpenHands within AgentCompass · 79 self-contained text-only tasks · three-run average · reasoning tiers and task-source commit not stated
Vals · With Skills · Accuracy
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · With Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
Vals · With Skills · Mean cost / test
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · With Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
Vals · With Skills · Mean duration / test
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · With Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
Vals · Without Skills · Accuracy
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · Without Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
Show 2 more resultsHide 2 results
Vals · Without Skills · Mean cost / test
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · Without Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
Vals · Without Skills · Mean duration / test
Reported by Vals AI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Vals AI · updated 2026-09-27 · metadata version 1 · 87 tasks in trace charts · 3 trials/task/condition · per-task mean then task average · OpenHands-CLI@ · Without Skills · system=zai/glm-5.3 · provider=Zhipu AI · reasoning_effort=max · temperature=1 · top_p=0.95 · max_output_tokens=128000
SuperCLUESource: SuperCLUE data
Vals17 related benchmarks · 19 results
Show 13 more resultsHide 13 results
Immigration
Source: Vals AI
LegalBench
Source: Vals AI
MedCode
Source: Clinical Benchmarks data docs · Sep 2026
Setting not stated
Source: Vals AI
Setting not stated
MedScribe
Source: Clinical Benchmarks data docs · Sep 2026
Setting not stated
Source: Vals AI
Setting not stated
Router Index
Source: Vals AI
SRE Bench
Source: Vals AI
TaxEval v2
Source: Vals AI
WideSearch · Atria paper scoreAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · in-house search/visit/python harness · 500 steps · 256K context / 64K max output · score subtype unstated · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Workspace-Bench · English Full · Atria runAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · AgentCompass · English Full · two-hour limit · 2 CPUs / 8 GiB · Seed 2.0 Lite judge · GPT in Codex, others in Claude Code · metric subtype unstated · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Workspace-Bench · English Lite · Atria runAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · AgentCompass · English Lite · two-hour limit · 2 CPUs / 8 GiB · Seed 2.0 Lite judge · GPT in Codex, others in Claude Code · metric subtype unstated · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
τ³-Bench · Banking Knowledge · Atria runAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · 97 Banking Knowledge tasks · modified Tau2/Tau3 runner from commit a2c0247 · alltools BM25/dense/shell retrieval · GPT-5.2 low judge and simulator · 200-step limit · Reported by Shanghai AI Laboratory · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats · SEA-HELM leaderboard, AI Singapore (AI Products team), National University of Singapore. Published aggregate scores; results licence not stated (SEA-HELM code is MIT). · Epoch AI Benchmarking Hub, CC BY 4.0 (licence) · Clinical Benchmarks (clinicalbenchmarks.ai), CC BY 4.0; fractions converted to percentages; source documents retain their own terms (licence) · LMArena, CC BY 4.0 (licence) · Bangla LLM Bench, Md Deluair Hossen. Source tasks: Hendrycks et al. 2021 (MMLU), OpenAI MMMLU, Bandarkar et al. 2024 (Belebele), Shi et al. 2023 (MGSM), Nahin et al. 2025 (TituLLMs). Public numeric community results; results licence not separately stated. · Baromètre de l’écriture IA, Johannes Eva; public numeric ranking data, © 2026 Johannes Eva, no results licence stated · Benchmark Nusantara, ibahasa (M. Khoirul Huda), doi:10.5281/zenodo.21970463. Test results CC BY 4.0, attribution to ibahasa; aggregate results CSV only. (licence) · BullshitBench by Peter Gostev, MIT (licence) · EQ-Bench (Samuel J. Paech) — cite paech2023eqbench · CyberGym leaderboard (Wang, Shi, He, Cai, Zhang, Song; UC Berkeley), ICLR 2026. Results evaluated and submitted by the listed teams. No licence is stated for the result data; the repository's Apache-2.0 licence covers its code only. · Keelemudelite mõõdupuu: Independent LLM leaderboard for Estonian. Eesti Keele Instituut. https://moodupuu.eki.ee, CC BY 4.0. Latest per-benchmark results selected with partial-run flags retained. (licence) · ExploitGym leaderboard (Wang et al., UC Berkeley), arXiv:2605.11086. No licence is stated for the result data; the repository's Apache-2.0 licence covers its code only, and task data keeps its upstream licences. Literal exploit counts; hidden rows omitted. · OpenRouter public benchmark leaderboard and search-lane exports, CC BY 4.0. Fractional source scores converted to percentages; immutable dated snapshots. (licence) · LLM Chess by Maxim Saplin, Apache-2.0 (Copyright [2024] [Maxim Saplin]) (licence) · MineBench, MIT (licence) · Outback Yak Research, NepNewsBench v2 research draft. Public numerical aggregates from a candidate preview that the publisher says is not evidence of publication; no release licence assigned and none asserted here. · Hugging Face official benchmark leaderboards; licence varies by benchmark · UzLiB, Tahrirchi (Abror Shopulatov). Public LEADERBOARD.md numbers; repository MIT covers software and documentation, no separate results licence stated and none asserted.

Also known as: @cf/zai-org/glm-5.3, accounts/fireworks/models/glm-5p3, accounts/fireworks/routers/glm-5p3-fast, aihubmix/coding-glm-5.3, aihubmix/glm-5.3, alibaba/glm-5.3, azure_ai/fw-glm-5.3, baseten/zai-org/glm-5.3, baseten/zai-org/glm-5.3-fast, cline-pass/glm-5.3, cloudflare/glm-5.3, coding-glm-5.3, coding-glm-5.3-free, databricks/databricks-glm-5-3, e2ee-glm-5-3-p, empiriolabs/glm-5-3, fireworks/glm-5.3, fireworks_ai/accounts/fireworks/models/glm-5p3, fireworks_ai/accounts/fireworks/routers/glm-5p3-fast, fireworks_ai/accounts/fireworks/routers/glm-5p3-us, fireworks_ai/glm-5p3, fireworks_ai/glm-5p3-fast, fireworks_ai/glm-5p3-us, glm 5.3, glm-5-3, glm-5.3-fast, glm-5.3-flex, glm-5.3-max, glm-5.3:fast, glm-5.3@eu, glm5.3, global.zai.glm-5.3, gmi/zai-org/glm-5.3, hf:zai-org/glm-5.3, mistral/glm-5.3, mistral/zai-glm-5-3, nebius/zai-org/glm-5.3, novita/glm-5.3, novita/zai-org/glm-5.3, opencode_go/glm-5.3, perplexity/glm-5.3, runpod/glm-5.3, runware/glm-5.3, scx-ai-gp/glm-5.3, scx-ai-gp/glm-5.3-fast, scx-ai/glm-5.3, streamlake/glm-5.3, tee/glm-5.3, tencent/glm-5.3, together-ai/glm-5.3, together_ai/zai-org/glm-5.3, umans-glm-5.3, us.zai.glm-5.3, wandb/zai-org/glm-5.3, z-ai-glm-5-3, z-ai-glm-5.3, z-ai/glm-5.3, z-ai/glm-5.3-fast, z-ai/glm-5.3-free, z-ai/glm-5.3:batch, z-ai/glm-5.3:thinking, zai-glm-5-3, zai-org/glm-5.3, zai-org/glm-5.3-bf16, zai-org/glm-5.3-fast, zai-org/glm-5.3-fp8, zai.glm-5.3, zai/glm-5.3, zai/glm-5.3-fast, zhipu/glm-5-3, zhipu/glm-5.3

Like this one. Tap or hover any dotted value, label or icon to see where it comes from and how it is measured. open details when tapped or hovered. How this data works