Zhipu AI Open weights: zai-org/GLM-5.3Reasoning model

GLM 5.3

Reasoning effort

GLM 5.3 is a Zhipu AI open-weights reasoning model, released 18 Aug 2026. Its best configuration (max reasoning effort) ranks #55 of 756 on the BenchLeader Index at 64.1 ±3.3, in the upper half. The ± is the point: 118 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in composite (81) and lowest in knowledge (59). At $2.15 per million tokens blended it is pricier than most ranked models. Output speed of 71 tokens per second puts it slower than most, with a first answer in 31.7 s. It has been measured at 4 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 2 Oct 2026.

Blended price: $2.15/M; $1.40 in · $4.40 out
Output speed: 71 tok/s; measured by Artificial Analysis
First answer: 32 s; first token 1.08 s
Context: 1M
Full answer: 39 s; median, reasoning included
Cost per run: $2.01; one full Intelligence Index run
Released: 18 Aug 2026

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index64
Reasoning69
Coding62
Agents & tools61
Maths61
Knowledge59
Human preference68
Long context65
Composite81

Versions

Zhipu AI has shipped 8 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.

Model	Released	Index	Rank
GLM 5.3maxthis page	18 Aug 2026	64.1	#55
GLM 5.2max	13 Jun 2026	62.7	#74
GLM 5.1thinking	7 Apr 2026	56.8	#179
GLM-5thinking	11 Feb 2026	59.0	#129
GLM 4.7thinking	22 Dec 2025	52.2	#309
GLM-4.6	30 Sept 2025	50.0	#360
GLM-4.5	28 Jul 2025	51.9	#317
Glm 4 0520	–	–	–

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
low	56.0	#202	67 tok/s	33 s	$0.0019	Coding 44 · Composite 69 · Knowledge 58 · Long context 62 · Reasoning 66
high	–	–	77 tok/s	1.08 s	$0.0019
maxbest	64.1	#55	71 tok/s	32 s	$0.0019	Agents & tools 61 · Coding 62 · Composite 81 · Human preference 68 · Knowledge 59 · Long context 65 · Maths 61 · Reasoning 69
not stated	–	–	77 tok/s	1.08 s	$0.0019	Agents & tools 62 · Composite 53

What more thinking costs

Turning the effort up buys index points and multiplies the bill. Here max costs 2.4× low for +8.1 on the index — 6.6 points per doubling of spend. Compare that with other models.

How its index has moved

8 Sept 2026 to 5 Oct 2026

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Metadata

Benchmark	low	high	max	not stated	Source	Trend
AA Openness Indexnot in index	–	–	33.3#32	–	Artificial Analysis

Reasoning

Benchmark	low	high	max	not stated	Source
GPQA Diamond	–	–	90.9%#33	–	Epoch AI Benchmarking Hub
LMArena Hard Prompts	–	–	1503#25	–	LMArena
LiveBench Reasoningnot in index	–	–	–	85.8%#34	LiveBench
GPQA Diamond (AA)not in index	–	–	91.7%#45	–	Artificial Analysis
Humanity's Last Exam (AA)not in index	36.6%#99	–	42.3%#64	–	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	88.1%#34	–	Vals AI
CritPt	14.6%#86	–	19.1%#60	–	Artificial Analysis
MysteryMechanismnot in index	–	–	23.0%#15	–	Vals AI
Chess Puzzlesnot in index	–	–	21.0%#79	–	Epoch AI Benchmarking Hub
Mystery Game Puzzlesnot in index	–	–	33.0%#24	–	Epoch AI Benchmarking Hub
LMCAnot in index	–	–	–	55.5%#31	Epoch AI Benchmarking Hub
DTBenchnot in index	–	–	–	87.7%#83	Epoch AI Benchmarking Hub
BTF-3not in index	–	–	–	14.9%#1	Epoch AI Benchmarking Hub

Coding

Benchmark	low	high	max	not stated	Source
SciCode	42.0%#143	–	59.0%#14	–	SciCode
WeirdML	–	–	75.4%#29	–	WeirdML
FrontierCode	–	–	40.1%#23	–	Cognition
LMArena Coding	–	–	1521#34	–	LMArena
LMArena WebDev	–	–	1623#19	–	LMArena
LiveBench Codingnot in index	–	–	–	79.0%#25	LiveBench
SciCode (AA)not in index	42.0%#145	–	59.0%#14	–	Artificial Analysis
LiveCodeBench	–	–	80.5%#65	–	Vals AI
IOI	–	–	68.4%#12	–	Vals AI
SWE-bench (Vals)not in index	–	–	95.4%#5	–	Vals AI
Code Migrationnot in index	–	–	44.2%#20	–	Vals AI
ProgramBenchnot in index	–	–	1.5%#9	–	Vals AI
Vibe Code Bench 1-100not in index	–	–	20.0%#8	–	Vals AI
Vibe Code Bench v1.1not in index	–	–	78.1%#24	–	Vals AI
DeepSWEnot in index	–	–	69.0%#16	–	Epoch AI Benchmarking Hub
CursorBenchnot in index	33.3%#43	38.0%#30	42.6%#20	–	Cursor
ALE-Benchnot in index	–	1317.4#26	–	–	Epoch AI Benchmarking Hub
FrontierSWEnot in index	–	–	30.2%#9	–	Epoch AI Benchmarking Hub
Terminal-Bench 4.0 (AA)not in index	34.9%#32	–	41.9%#26	–	Artificial Analysis
Terminal-Bench 2.1 (AA)not in index	–	–	83.9%#34	–	Artificial Analysis

Agents & tools

Benchmark	low	high	max	not stated	Source
Terminal-Bench	–	–	41.8%#43	–	Terminal-Bench
APEX-Agents	–	–	–	56.6%#17	Mercor
LMArena Agent	–	–	2.4#22	–	LMArena
LiveBench Agentic Codingnot in index	–	–	–	60.9%#14	LiveBench
Terminal-Bench 2.1 (Vals)	–	–	71.5%#26	–	Vals AI
MCP Atlas	–	–	–	84.2%#5	Scale AI SEAL
GDPval (AA)not in index	40.2%#88	–	57.8%#15	–	Artificial Analysis
τ²-Bench Banking (AA)	–	–	50.3%#4	–	Artificial Analysis
ITBench SRE (AA)not in index	–	–	46.1%#12	–	Artificial Analysis
CyberBenchnot in index	–	–	72.1%#14	–	Vals AI
Finance Agent v2not in index	–	–	55.8%#18	–	Vals AI
Harvey's Legal Agent Benchmarknot in index	–	–	8.3%#19	–	Vals AI
Legal Research Benchnot in index	–	–	49.0%#7	–	Vals AI
SkillsBenchnot in index	–	–	47.5%#24	–	Vals AI
Tax Agent Benchnot in index	–	–	73.1%#5	–	Vals AI
Terminal-Bench 4.0 (Vals)not in index	–	–	38.9%#10	–	Vals AI
Terminal-Bench Sciencenot in index	–	–	4.3%#17	–	Vals AI
Vending-Bench 2not in index	–	–	–	8163.6#11	Andon Labs
AutomationBenchnot in index	–	–	62.2%#18	–	Artificial Analysis
GDP.pdfnot in index	–	–	11.2%#39	–	Artificial Analysis
EnterpriseOps-Gymnot in index	–	–	36.4%#11	–	Artificial Analysis
AA-Briefcasenot in index	–	–	1510#22	–	Artificial Analysis

Maths

Benchmark	low	high	max	not stated	Source
FrontierMath Tiers 1–3	–	–	68.8%#25	–	Epoch AI Benchmarking Hub
FrontierMath Tier 4	–	–	29.3%#35	–	Epoch AI Benchmarking Hub
OTIS Mock AIME	–	–	91.1%#62	–	Epoch AI Benchmarking Hub
ProofBench	–	–	49.0%#31	–	Vals AI
LiveBench Mathematicsnot in index	–	–	–	87.9%#41	LiveBench
LMArena Maths	–	–	1494#20	–	LMArena

Knowledge

Benchmark	low	high	max	not stated	Source
SimpleQA Verified	–	–	41.0%#42	–	Epoch AI Benchmarking Hub
LiveBench Data Analysisnot in index	–	–	–	70.2%#48	LiveBench
AA-Omniscience	-8.4#165	–	14.3#80	–	Artificial Analysis
MMLU-Pro	–	–	86.8%#37	–	Vals AI
LegalBench	–	–	84.8%#27	–	Vals AI
TaxEval	–	–	72.4%#64	–	Vals AI
Excel Modeling Benchmarknot in index	–	–	56.3%#40	–	Vals AI
MedCodenot in index	–	–	42.9%#46	–	Vals AI
MedScribenot in index	–	–	88.8%#9	–	Vals AI
Public Benefits Benchnot in index	–	–	68.5%#7	–	Vals AI
AA-Omniscience: accuracynot in index	33.9%#149	–	33.9%#150	–	Artificial Analysis
AA-Omniscience: non-hallucinationnot in index	36.0%#170	–	70.5%#37	–	Artificial Analysis

Instruction following

Benchmark	low	high	max	not stated	Source
LiveBench Languagenot in index	–	–	–	79.9%#30	LiveBench
LiveBench Instruction Followingnot in index	–	–	–	69.3%#31	LiveBench
LMArena Instruction Followingnot in index	–	–	1476#25	–	LMArena

Human preference

Benchmark	low	high	max	not stated	Source
LMArena Text	–	–	1479#27	–	LMArena
LMArena Creative Writingnot in index	–	–	1454#34	–	LMArena
LMArena Multi-turnnot in index	–	–	1480#41	–	LMArena
LMArena Longer Queriesnot in index	–	–	1486#34	–	LMArena

Long context

Benchmark	low	high	max	not stated	Source	Trend
AA-LCR	72.7%#173	–	79.7%#80	–	Artificial Analysis
MLCRnot in index	–	–	48.3%#10	–	Artificial Analysis

Composite

Benchmark	low	high	max	not stated	Source
Epoch Capabilities Indexnot in index	–	–	–	155.8#24	Epoch AI Benchmarking Hub
LiveBench	–	–	–	76.1%#31	LiveBench
AA Intelligence Index	34.3#83	–	44.8#36	–	Artificial Analysis
Vals Indexnot in index	–	–	53.5#16	–	Vals AI

Safety & honesty

Benchmark	low	high	max	not stated	Source	Trend
FORTRESSnot in index	–	–	28.2%#26	–	Scale AI SEAL

Where it wins

Benchmarks where this configuration ranks in the top five of every configuration measured.

#4τ²-Bench Banking (AA)50.3% · of 201

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
Mistral	146 tok/s	0.87 s	$1.54	$4.84	1.0M	nvfp4
Crusoe	136 tok/s	0.65 s	$1.40	$4.40	1.0M	fp4
Baseten Fast	136 tok/s	1.13 s	$2.10	$6.60	1.0M	fp8
Decart	133 tok/s	0.52 s	$1.19	$3.74	1.0M	fp4
Mistral	131 tok/s	1.19 s	$1.40	$4.40	1.0M	nvfp4
Prime Intellect	129 tok/s	1.16 s	$1.40	$4.40	1.0M	–
Mistral	123 tok/s	1.22 s	$1.40	$4.40	1.0M	nvfp4
Nebius Token Factory	122 tok/s	1.08 s	$1.40	$4.40	1.0M	fp4
Fireworks Fast	122 tok/s	1.16 s	$2.10	$6.60	1.0M	–
Wafer	121 tok/s	0.75 s	$0.100	$7.00	1.0M	–
Together	112 tok/s	0.62 s	$1.40	$4.40	1.0M	–
Wafer	110 tok/s	0.94 s	$0.050	$7.00	1.0M	–
Sail Research	110 tok/s	1.06 s	$0.200	$3.40	1.0M	fp8
Friendli	98 tok/s	1.93 s	$1.26	$3.96	1.0M	–
Baseten Fast	97 tok/s	1.03 s	$2.10	$6.60	1.0M	fp8
Modal	96 tok/s	0.81 s	$1.40	$4.40	1.0M	–
Parasail	94 tok/s	0.75 s	$1.40	$4.40	1.0M	fp8
inference.net	91 tok/s	0.70 s	$0.120	$4.40	1.0M	–
Sail Research	90 tok/s	1.47 s	$0.200	$3.40	1.0M	fp8
AkashML	87 tok/s	1.55 s	$1.05	$3.56	1.0M	fp8
SiliconFlow	76 tok/s	1.16 s	$0.700	$2.20	1.0M	fp8
Relace	72 tok/s	1.24 s	$0.030	$12.00	1.0M	–
Makora	71 tok/s	1.19 s	$0.240	$3.93	980k	fp4
AtlasCloud	68 tok/s	1.28 s	$1.40	$4.40	1.0M	fp8
Baidu Qianfan	63 tok/s	0.76 s	$1.40	$4.40	1.0M	fp8
Alibaba Prime	62 tok/s	1.17 s	$2.80	$8.80	1M	–
Alibaba Cloud Int.	60 tok/s	1.23 s	$1.19	$3.74	1M	–
Baseten	59 tok/s	5.32 s	$1.40	$4.40	1.0M	fp4
GMICloud	57 tok/s	1.99 s	$0.980	$3.08	1.0M	fp8
Baseten	53 tok/s	5.89 s	$1.40	$4.40	1.0M	fp4
Z.ai	50 tok/s	3.07 s	$1.40	$4.40	1.0M	fp8
DigitalOcean	45 tok/s	0.89 s	$0.910	$2.86	1.0M	–
Reka AI	44 tok/s	1.52 s	$0.170	$3.00	262k	–
Fireworks	43 tok/s	1.02 s	$1.40	$4.40	1.0M	–
Inceptron	41 tok/s	0.92 s	$0.600	$3.39	1.0M	fp4
Cloudflare	40 tok/s	5.07 s	$1.40	$4.40	1.0M	–
Phala	38 tok/s	4.09 s	$0.840	$2.64	1.0M	–
DeepInfra	37 tok/s	1.75 s	$0.563	$2.50	1.0M	fp4
Venice	35 tok/s	2.83 s	$1.40	$4.40	1M	–
Morph	27 tok/s	1.16 s	$0.131	$3.18	1.0M	fp8
NovitaAI	26 tok/s	7.46 s	$0.420	$1.32	1.0M	fp8

Price history

Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.

input outputnow $0.600 in · $3.39 out

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.260 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0019	$0.0015	35.9 s
Summarise a 30-page report	12,000 / 600	$0.019	$0.0092	40.2 s
Code edit	6,000 / 1,500	$0.015	$0.0099	52.9 s
Agentic coding session	60,000 / 4,000	$0.102	$0.050	1.5 min
Structured extraction	2,000 / 200	$0.0037	$0.0020	34.5 s