GPT-5.2

+ contribute

GPT-5.2 is a closed-weights language model from OpenAI, released December 2025.

73.8±0.5#51 of 719 LLMsfrom 50 benchmark results across 26 sources

$ how it compares

The LLMs ranked either side of GPT-5.2, out of 828 with a score, then more from OpenAI.

#	model	score	where	RAM	released
52	GLM-5.3-Flash Z.ai	74±1	api $0.24/M +local	~103GB	2026-08
53	Gemini 3.5 Flash Google	74±1	api $3.38/M		2026-05
54	GPT-5.2 OpenAI	74±1	api $4.81/M		2025-12
55	Hunyuan Hy3 (295B-A21B) Tencent	74±1	api $0.23/M +local	~96GB	2026-07
56	Gemini 3.6 Flash Google	74±1	api $1.50/M		2026-07
	more from OpenAI
1	GPT-6 Astra OpenAI	87±1	api $20.00/M		2026-09
3	GPT-6.1 Sol OpenAI	85±1	api $4.00/M		2026-09
7	GPT-6 Sol OpenAI	83±1	api $4.00/M		2026-09

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #58 of 300 · 23% of its score · 14 tests

Beats Gemini 3.1 Pro (Preview) here, though they rank higher overall. Trails MiniMax-M3, GLM-5.1, Claude Opus 4.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified	73.8% ✎	▁··▁·▁▂▁▄█▄▁ beats 14/29	hard · avg 19%	▇▇ 4.4%	sharp
MCP-Atlas (public set) +1	68.0% ✎	▄······▄▄▄▄█ beats 6/6	hard · avg 25%	▇ 3.2%	at its level · several runs
Terminal-Bench 2.0 with Terminus 2 agent (benchmark authors' runs)	54.0% ✎	▃··▃·▃▃██▅▃▅ beats 13/14	very hard · avg 5.8%	▇ 2.7%	sharp · at its level
ContractBench success rate (33 observation-contract tool tasks)	74.7% ✎	▆··▆▄▂▄▂▆▆█▄ beats 22/23	very hard · avg 14%	▇ 1.9%	sharp · at its level
τ³-bench banking_knowledge (v1.0.1 grading)	32.2% ✎	▃▃▂▃·▆▃▃█▅▅▂ beats 13/25	very hard · avg 4.1%	▇ 1.8%	at its level
τ²-bench aggregate	87.1% ✎	▄····█▄▄···▄ beats 4/5	medium · avg 40%	▇ 1.5%	typical
Vibe Code Bench (test split, 50 web apps)	53.5% ✎	▃··▃·▃██▃▃▃█ beats 13/14	very hard · avg 2.3%	▇ 1.4%	sharp · at its level · noisy
Thinkingbox-bench pass@1 (507 stateful business-workflow agent tasks, N=20)	46.3% ✎	▅··▃·▃·█▃▅██ beats 8/15	very hard · avg 4.9%	▇ 1.2%	at its level

6 more tests

test	result	vs the others in its run	difficulty	counts	why
BFCL V4 (December 2025 archive)	55.9% ✎	▁▁▁▂██▅▅▆▂▂▂ beats 66/77	hard · avg 32%	▇ 1.1%	at its level
τ³-bench airline (tau2-bench ≥ v1.0.1)	83.0% ✎	▃······▃·▃█▃ beats 5/6	medium · avg 49%	▇ 1.1%	typical
τ³-bench retail (tau2-bench ≥ v1.0.1)	81.6% ✎	▄▄·█··▄·▄··▄ beats 5/6	medium · avg 46%	▇ 1.0%	typical
τ³-bench telecom (tau2-bench ≥ v1.0.1)	89.7% ✎	█·▄█▄······▄ beats 2/6	medium · avg 62%	▇ 0.8%	maxed out
τ²-bench Telecom lab	98.7% ✎	▃·▅▅▃▃··█▃▃▅ beats 13/13	hard · avg 16%	▇ 0.7%	sharp · noisy
τ²-Bench-Verified (GPT-5.1 user)	80.3% ✎	▂··▄···▂▄▄█▄ beats 11/13	medium · avg 45%	▇ 0.4%	typical

coding▁▁▁▃▆██▇▆▄▃▁ #79 of 458 · 10% of its score · 3 tests

Trails Llama 3.1 Nemotron Ultra 253B v1, GLM-4.7, Gemini 3.5 Flash-Lite here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1415.5 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 49/119	avg 981	▇▇ 4.4%	sharp
SWE-bench Verified	73.8% ✎	▁··▁·▁▂▁▄█▄▁ beats 14/29	hard · avg 19%	▇▇ 4.4%	sharp
Vibe Code Bench (test split, 50 web apps)	53.5% ✎	▃··▃·▃██▃▃▃█ beats 13/14	very hard · avg 2.3%	▇ 1.4%	sharp · at its level · noisy

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #70 of 393 * · 2.5% of its score · 1 test

Trails Gemma 4 26B-A4B, Nemotron Cascade 2 30B-A3B, Gemma 4 31B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1437.3 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 294/379	avg 1268	▇ 2.5%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #31 of 398 · 31% of its score · 9 tests

Beats DeepSeek V4.1 Flash, DeepSeek V4 Pro (0813), Claude Opus 4.7 here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tiers 1–3 v2	67.4% ✎	▂··▁▁▃▄▄█▅▃▃ beats 56/82	very hard · avg 1.2%	▇▇▇▇▇ 12%	sharp · at its level
FrontierMath tier 4 v2	31.7% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 36/63	very hard · avg 0.1%	▇▇▇ 8.6%	sharp
MathArena ArXivMath February 2026 (32 problems)	37.5% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 10/24	very hard · avg 9.1%	▇ 3.2%	at its level
MathArena ArXivMath January 2026 (23 problems)	67.9% ✎	▂··▄▂▄█▆▆▄▄█ beats 17/23	hard · avg 22%	▇ 2.1%	at its level
MathArena ArXivMath December 2025 (17 problems)	52.2% ✎	▃·▃▃█▁·▄·▁▁▁ beats 15/18	hard · avg 19%	▇ 1.5%	at its level
MathArena BrokenArXiv February 2026 (31 problems)	25.8% ✎	▃▂▂▃█▃▃·▃··▂ beats 13/16	very hard · avg 0.7%	▇ 1.2%	sharp · noisy
MATH-Vision	83.0% ✎	█·····▄▄·▄·▄ beats 2/5	medium · avg 46%	▇ 0.9%	typical
OTIS mock AIME 2024–2025	96.1% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 166/193	very hard · avg 8.5%	▇ 0.6%	sharp · maxed out · noisy

1 more test

test	result	vs the others in its run	difficulty	counts	why
FormalProofBench (200 graduate-level Lean 4 problems, agentic)	13.5% ✎	▃▅·▃··▅▃█··▃ beats 5/10	very hard · avg 3.0%	▇ 0.6%	noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #44 of 147 · 4.9% of its score · 5 tests

Beats MiMo-V2.6-Pro, MiMo-V2.6-Flash (309B-A15B), DeepSeek V4 Flash Vision (Exp) here, though they rank higher overall. Trails Qwen3.6 Flash, Gemini 3.5 Flash-Lite, Kimi K2.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1242.8 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 95/145	avg 1068	▇ 1.7%	blunt
CharXiv (reasoning questions) lab	82.1% ✎	▄··▂█··▂···▄ beats 9/9	hard · avg 34%	▇ 1.2%	typical
MATH-Vision	83.0% ✎	█·····▄▄·▄·▄ beats 2/5	medium · avg 46%	▇ 0.9%	typical
Arena document (style controlled)	1423.2 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 3/37	avg 1335	▇ 0.8%	blunt
Diagram-MMU diagram question answering (DQA accuracy, scientific TikZ diagrams)	81.7% ✎	▄·····▄▄·█·█ beats 2/6	medium · avg 55%	▇ 0.4%	blunt

web search▂▁▂▆▄█▃▃▁▁·▁ #51 of 80 * · 1.0% of its score · 1 test

Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena search systems (style controlled)	1161.5 ✎	▁·▁▁▂▃▂█▅▂▁▁ beats 7/30	avg 1055	▇ 1.0%	blunt

other tests19 not in a task · 34% of its score

test	result	vs the others in its run	difficulty	counts	why
EBR-bench (Epoch)	23.0% ✎	▄▂·▆█▂▆▄▆▆·▄ beats 11/23	very hard · avg 0.3%	▇▇ 4.0%	sharp
Furniture Assembly (Epoch)	38.3% ✎	█▆▆▆▆▂▅▂▂▂·▅ beats 15/30	very hard · avg 1.2%	▇▇ 3.8%	sharp · at its level
KMMLU-HARD (Daekeun Kim, zero-shot)	74.6% ✎	█···▁▁▁▂▃·▆▂ beats 18/23	very hard · avg 29%	▇ 3.3%	sharp · at its level
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)	27.6% ✎	█··▂▅▁▃▃▃·▁▃ beats 17/24	very hard · avg 0.3%	▇ 3.0%	sharp · at its level · noisy
Mystery Game Puzzles (Epoch)	23.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 41/74	very hard · avg 9.4%	▇ 2.8%	sharp · noisy
GPQA Diamond	91.4% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 184/208	hard · avg 40%	▇ 2.2%	sharp
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	39.5% ✎	▁▁▂▁▃▄█▆▄▅▂▃ beats 33/39	very hard · avg 22%	▇ 1.9%	blunt · at its level
Chess Puzzles (Epoch)	49.0% ✎	█▁▁▁▁▂▁▄▂▂▃▁ beats 127/135	very hard · avg 5.9%	▇ 1.9%	sharp · noisy

11 more tests

test	result	vs the others in its run	difficulty	counts	why
Medmarks-V weighted mean win rate (61 models, 71 configurations)	62.4% ✎	▂▃▅▂▆▃█▅█▅▃▆ beats 32/34	medium · avg 42%	▇ 1.6%	blunt · at its level
EQ-Bench 3 (emotional intelligence, judge Elo)	1314.3 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 53/62	avg 1044	▇ 1.5%	typical
CLIcK Korean (Daekeun Kim)	90.9% ✎	▆▄▄▄█▄▂·▄█·▄ beats 18/23	medium · avg 58%	▇ 1.4%	typical
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	95.0% ✎	▁·▇▃▁·█▇▁·▇▃ beats 22/27	very hard · avg 4.1%	▇ 1.3%	sharp · noisy
RPC-Bench overall, text input (Z.ai)	35.2% ✎	▄··▄▄·▄▄█·▄▄ beats 7/8	very hard · avg 11%	▇ 1.2%	at its level
HAE-RAE 1.0 (Daekeun Kim)	91.8% ✎	▄▄▄▂▂█▂▄▂█▂▄ beats 16/22	medium · avg 63%	▇ 1.0%	typical
SimpleQA Verified	37.1% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 32/78	very hard · avg 8.0%	▇ 1.0%	at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)	47.0% ✎	▁▁▃▂▂▄▁▃▄▄▄█ beats 18/38	very hard · avg 14%	▇ 0.9%	at its level
HLE-Verified full set, text-only (Qwen Team runs)	43.3% ✎	█·█·····█··█ beats 2/3	very hard · avg 14%	▇ 0.8%	at its level
FinSheet-Bench accuracy (501 questions, 24 spreadsheets)	80.4% ✎	█···██·····█ beats 2/3	medium · avg 52%	▇ 0.6%	blunt
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	89.2% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 35/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $4.81/M · $1.75 in / $14 out per 1M tokens · 400k context · platform.openai.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

$ how the score was worked out · Epoch AI — internal evaluatio… 41%, LMArena leaderboard dataset 10%, MathArena ArXivMath / BrokenA… 8% and 23 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	10	▇▇▇▇▇▇▇▇▇▇▇▇ 41%	CC BY 4.0
LMArena leaderboard dataset	5	▇▇▇ 10%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	4	▇▇ 8%	CC BY-SA 4.0
Daekeun Kim Korean LLM benchmark runs	3	▇▇ 6%	MIT
τ²/τ³-bench leaderboard: Sierra's own evaluation runs	4	▇ 5%	MIT
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	2	▇ 4%	MIT
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces — benchmark authors' own agent-model trials	1	▇ 3%	CC BY 4.0
Qwen/Qwen3.5-397B-A17B model card	2	▇ 2%	apache-2.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 2%	CC BY 4.0
ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)	1	▇ 2%	CC BY 4.0
OpenAI vendor-reported benchmark scores for its own models (openai.com announcement posts and Deployment Safety Hub system cards) · lab-reported	2	▇ 2%	Factual vendor-reported score (not licensed content)
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models	1	▇ 2%	CC BY 4.0
zai-org/GLM-5 model card	1	▇ 2%	MIT
GLM-5 Team (Zhipu AI) — GLM-5 technical report	1	▇ 2%	CC BY 4.0
EQ-Bench 3 canonical leaderboard results	1	▇ 2%	MIT
Vibe Code Bench (Vals AI, arXiv:2603.04601v3)	1	▇ 1%	CC-BY-4.0
RPC-Bench leaderboard (Z.ai research paper comprehension)	1	▇ 1%	Apache-2.0
One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)	1	▇ 1%	CC BY 4.0
Berkeley Function Calling Leaderboard result archive	1	▇ 1%	Apache-2.0
Miðeind Icelandic LLM Leaderboard	1	▇ 1%	CC BY 4.0
HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam — Qwen Team re-evaluation on revised Humanity's Last Exam	1	▇ 1%	CC BY 4.0
FinSheet-Bench: From Simple Lookups to Complex Reasoning, Where LLMs Break on Financial Spreadsheets — authors' own model runs	1	▇ 1%	CC BY 4.0
FormalProofBench (Vals AI, arXiv:2603.26996v1)	1	▇ 1%	CC-BY-4.0
Amazon AGI — τ²-Bench-Verified	1	▇ 0%	MIT
Diagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams (arXiv:2608.12262v1)	1	▇ 0%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 73.8