GPT-6 Astra Medium

Scores

Dimension	Score	Rank	Benchmarks
Overall: Balanced	64.1	31	8
Overall: Coding	62.1	60	7
Overall: Reasoning	68.3	14	5
Overall: Agentic work	63.1	40	7
Overall: Frontend	60.8	61	2
Coding	60.8	73	2
Agents	62.9	25	1
Reasoning	67.5	21	4
Math	69.3	6	1
Value vs price	+5.4	127

Unverified rumours 😃 Physical AI anxiety and Huawei Ascend pretraining: Sources close to DeepSeek note that OpenAI's GPT 6 Astra hitting 95% accuracy on a robotic control task (versus 40% for Claude Fable 5.1) triggered internal concern that scaling text and code LLMs alone wi…@IamEmily2050 · 28 likes · negative
TOWARD WEIGHT-SPACE PROGRAMMING MIT CSAIL researchers recently had a frontier model, GPT-6 Astra, fix another AI by editing its weights directly. No training. They call it weight-space programming. Could a model running on my own Mac do that kind of brain surgery? Test 1 The…@NickGeorgeSEO · 5 likes
Which model has the best overall intelligence? - Gemini 4 Carbon - Fable 5.2 - Opus 5.5 - GPT-6.2 Sol - GPT-6 Astra@ItsKamranK · 2 likes

Benchmark	Result	Source	Date
DeepSWE	72.8%	Epoch AI	2026-11-11
SciCode	54.2%	Epoch AI	2026-11-11

Benchmark	Result	Source	Date
Terminal-Bench 4	55.2%	Terminal-Bench	2026-09-03

Benchmark	Result	Source	Date
CritPt	29.1%	Epoch AI	2026-11-11
ARC-AGI-2	92.1%	Epoch AI	2026-11-11
ARC-AGI-1 (ARC Prize verified)	97.5%	ARC Prize	2026-11-11
ARC-AGI-1	97.5%	Epoch AI	2026-11-11
ARC-AGI-2 (ARC Prize verified)	92.1%	ARC Prize	2026-11-11
ARC-AGI-3	98.4%	ARC Prize	2026-11-11

Benchmark	Result	Source	Date
FrontierMath Tier 4	97.6%	Epoch AI	2026-08-30

Benchmark	Result	Source	Why not used
ARC-AGI-3	38.6%	ARC Prize	Lower duplicate run on the same benchmark

gpt-6-astra-medium