GPT-6 Astra

+ contribute

GPT-6 Astra is a closed-weights language model from OpenAI, released September 2026.

87.4±0.6#1 of 719 LLMsfrom 39 benchmark results across 10 sources

$ how it compares

The LLMs ranked either side of GPT-6 Astra, out of 828 with a score, then more from OpenAI.

#	model	score	where	released
1	GPT-6 Astra OpenAI	87±1	api $20.00/M	2026-09
2	Claude Opus 5.5 Anthropic	87±1	api $8.00/M	2026-09
3	GPT-6.1 Sol OpenAI	85±1	api $4.00/M	2026-09
4	Claude Fable 5.1 Anthropic	85±1	api $20.00/M	2026-09
5	Claude Sonnet 5.5 Anthropic	84±1	api $4.00/M	2026-09
	more from OpenAI
7	GPT-6 Sol OpenAI	83±1	api $4.00/M	2026-09
8	GPT-5.5 Pro OpenAI	82±1	api $67.50/M	2026-04
10	GPT-5.6 Sol OpenAI	82±1	api $8.00/M	2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▃ #5 of 300 · 22% of its score · 8 tests

Trails Claude Sonnet 5.5, Claude Fable 5.1, GPT-6.1 Sol here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	0.1227 IPS ✎	▁▁▁▄▂▃▇█▂▂▁▃ beats 46/49	avg -0.312 IPS	▇▇▇ 7.5%	sharp
Terminal-Bench 4.0 lab	57.9% ✎	▁·▁··▁▇▇▅██▃ beats 27/30	very hard · avg 0.4%	▇▇ 4.1%	sharp · at its level · small run
Humanity's Last Exam with tools lab	57.2% ✎	▃·▃·▅····▅·█ beats 7/8	very hard · avg 12%	▇ 2.2%	at its level
BrowseComp lab	91.5% ✎	▄··█▄··▄·▄█▄ beats 8/8	hard · avg 32%	▇ 2.2%	typical
KlinikeBench strict pass@1 (333 interactive clinical encounters, Terminus-2)	24.3% ✎	▁▁·▁▁▁▁▅▃█▅█ beats 25/27	very hard · avg 2.4%	▇ 1.7%	at its level
DeepSWE v1.1 (resolved) lab	74.1% ✎	█··▃▃····▃·▃ beats 6/6	hard · avg 21%	▇ 1.6%	typical
Thinkingbox-bench pass@1 (507 stateful business-workflow agent tasks, N=20)	58.3% ✎	▅··▃·▃·█▃▅██ beats 11/15	very hard · avg 4.9%	▇ 1.4%	at its level
Agents' Last Exam (metric/harness unspecified) lab	59.3% ✎	▂·▂·█▂···▂▄▄ beats 11/11	very hard · avg 11%	▇ 1.4%	at its level

coding▁▁▁▃▆██▇▆▄▃▃ #2 of 458 · 9.4% of its score · 3 tests

Trails Claude Opus 5.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1787.7 ✎	▁▂▄▃▅█▄▇▆▄▁▃ beats 118/119	avg 981	▇▇▇ 6.7%	sharp
DeepSWE v1.1 (resolved) lab	74.1% ✎	█··▃▃····▃·▃ beats 6/6	hard · avg 21%	▇ 1.6%	typical
Luau Leetcode fully solved (Torpedo Software)	87.0% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 48/49	hard · avg 18%	▇ 1.1%	typical

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #23 of 393 * · 3.7% of its score · 1 test

Trails Gemini 3.6 Flash, Claude Opus 4.6, Muse Spark 1.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1477.1 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 354/379	avg 1268	▇ 3.7%	blunt

math▁▂▃▅▆▅▄█▄▁▁▃ #1 of 398 · 29% of its score · 11 tests

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tier 4 v2	97.6% ✎	▁▂▁▂█▅▂▂▁▁▃▁ beats 62/63	very hard · avg 0.1%	▇▇ 5.4%	sharp
MathArena ArXivMath April 2026 (40 problems)	90.8% ✎	▁·····▅██▄·▃ beats 20/20	very hard · avg 7.2%	▇▇ 5.0%	sharp
MathArena ArXivLean June 2026 (46 problems)	65.2% ✎	▄····▄▄███·▄ beats 9/9	very hard · avg 0.7%	▇▇ 3.9%	sharp · at its level
MathArena ArXivMath June 2026 (48 problems)	94.4% ✎	▂···▄·██▆▆▄▃ beats 19/19	very hard · avg 4.6%	▇▇ 3.8%	sharp
FrontierMath tiers 1–3 v2	93.7% ✎	▂··▁▁▃▄▄█▅▃▃ beats 82/82	very hard · avg 1.2%	▇ 3.6%	sharp · maxed out
MathArena BrokenArXiv April 2026 (61 problems)	97.5% ✎	▂▂▆█▃▃▂▂▂··▃ beats 18/18	very hard · avg 0.7%	▇ 2.8%	sharp · noisy
MathArena ArXivMath May 2026 (40 problems)	95.0% ✎	▁··▁▄█▄▁·▃·▃ beats 17/17	very hard · avg 10%	▇ 1.7%	typical
MathArena ArXivMath August 2026, with tools (57 problems)	88.6% ✎	█▄·▄·······█ beats 5/5	very hard · avg 4.8%	▇ 1.4%	sharp

3 more tests

test	result	vs the others in its run	difficulty	counts	why
FrontierMath: Erdős problems	2.9% ✎	█··········▆ beats 6/6	very hard · avg 0.0%	▇ 0.9%	sharp · too hard for it · noisy
MathArena BrokenArXiv August 2026, with tools (56 problems)	81.9% ✎	▄▄·····▄·▄·█ beats 5/5	very hard · avg 5.5%	▇ 0.7%	at its level · noisy
OTIS mock AIME 2024–2025	100.0% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 192/193	very hard · avg 8.5%	▇ 0.1%	sharp · maxed out · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #9 of 147 · 5.6% of its score · 5 tests

Trails Claude Opus 4.7, Muse Spark 1.2, Gemini 3.7 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1284.3 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 130/145	avg 1068	▇ 2.6%	blunt
Arena document (style controlled)	1472.6 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 25/37	avg 1335	▇ 1.2%	blunt
CharXiv (reasoning questions)	89.3% ✎	▂▅▃▅▅▆▆▅▆█▆▃ beats 100/100	hard · avg 33%	▇ 1.0%	small run
CharXiv descriptive questions	96.2% ✎	▂▄▄▅▇▄█▇▇▂▁▃ beats 58/58	easy · avg 66%	▇ 0.4%	maxed out · small run
ScreenSpot-Pro lab	92.7% ✎	▂▃▅▆██▅▂···▃ beats 24/24	medium · avg 48%	▇ 0.3%	blunt · small run

web search▂▁▂▆▄█▃▃▁▁·▁ #9 of 80 * · 4.4% of its score · 2 tests

Trails GPT-4o (2024-08-06), GPT-4.1 mini, GPT-4.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam with tools lab	57.2% ✎	▃·▃·▅····▅·█ beats 7/8	very hard · avg 12%	▇ 2.2%	at its level
BrowseComp lab	91.5% ✎	▄··█▄··▄·▄█▄ beats 8/8	hard · avg 32%	▇ 2.2%	typical

other tests12 not in a task · 31% of its score

test	result	vs the others in its run	difficulty	counts	why
EBR-bench (Epoch)	76.2% ✎	▄▂·▆█▂▆▄▆▆·▄ beats 23/23	very hard · avg 0.3%	▇▇▇▇ 10%	sharp · at its level
Mystery Game Puzzles (Epoch)	84.0% ✎	█··▂▃▇▄█▂▃▁▃ beats 74/74	very hard · avg 9.4%	▇▇ 6.0%	sharp · at its level · noisy
Furniture Assembly (Epoch)	80.0% ✎	█▆▆▆▆▂▅▂▂▂·▅ beats 29/30	very hard · avg 1.2%	▇▇ 5.5%	sharp · at its level
Chess Puzzles (Epoch)	72.0% ✎	█▁▁▁▁▂▁▄▂▂▁▃ beats 135/135	very hard · avg 5.9%	▇ 3.3%	sharp · at its level · noisy
SimpleQA Verified	75.6% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 78/78	very hard · avg 8.0%	▇ 1.2%	at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)	70.3% ✎	▁▁▃▂▂▄▁▃▄▄▄█ beats 37/38	very hard · avg 14%	▇ 1.1%	at its level
HealthBench Hard lab	34.2% ✎	▂··▂····▂▄·█ beats 7/8	very hard · avg 9.7%	▇ 1.0%	blunt · at its level
MirrorCode (Epoch)	46.7% ✎	█·▄▄·▄·▄·▄·█ beats 5/8	very hard · avg 1.5%	▇ 0.9%	sharp · at its level · noisy

4 more tests

test	result	vs the others in its run	difficulty	counts	why
HealthBench lab	56.9% ✎	▃·▃·····█▃▃▅ beats 6/8	hard · avg 24%	▇ 0.8%	blunt · at its level
GPQA Diamond	95.8% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 208/208	hard · avg 40%	▇ 0.7%	sharp · maxed out
RobloxQA v2.0 multiple choice (Torpedo Software)	98.5% ✎	▁▁▃▁█▄▃▃▆▂▂▃ beats 47/49	easy · avg 78%	▇ 0.4%	maxed out
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	91.3% ✎	▁▂▁▂▅█▃▃▄▃▁▁ beats 58/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $20.00/M · $10 in / $50 out per 1M tokens · 1.1M 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… 38%, LMArena leaderboard dataset 22%, MathArena ArXivMath / BrokenA… 19% and 7 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	11	▇▇▇▇▇▇▇▇▇▇▇ 38%	CC BY 4.0
LMArena leaderboard dataset	5	▇▇▇▇▇▇▇ 22%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	7	▇▇▇▇▇▇ 19%	CC BY-SA 4.0
OpenAI vendor-reported benchmark scores for its own models (openai.com announcement posts and Deployment Safety Hub system cards) · lab-reported	8	▇▇▇▇ 14%	Factual vendor-reported score (not licensed content)
KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (arXiv:2609.38480v2)	1	▇ 2%	CC BY 4.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇ 2%	MIT
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
Meridian Labs inspect_charxiv evaluation report (GitHub README)	2	▇ 1%	MIT
Miðeind Icelandic LLM Leaderboard	1	▇ 1%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 87.4