GPT-6 Astra
OpenAI · released
OpenAI's top tier, built for the hardest long-running agentic work: large software-engineering tasks, operating computers and browsers, research and polished deliverables. It is OpenAI's priciest model and the first it rates at its “Critical” cybersecurity level, so some security requests are refused.
Best for: Hardest agentic coding · Computer and browser use · Research and analysis
- Intelligence Index
- 52.7
- #4 of 38 · at Max effort
- Price per 1M tokens
- $10 in $50 out
- Blended $20 at 3:1
- Context window
- 1.05M
- Up to 128K output
- Output speed
- 48 tokens/s
- $3.26 measured cost per task
Capability profile
One 0–100 score per area, placing this model between the weakest and strongest tracked model across that area's benchmarks.
- Overall85 · #5 of 38
Based on 3 of 3 benchmarks
- Coding90 · #3 of 39
Based on 5 of 5 benchmarks
- Agents & knowledge work72 · #9 of 40
Based on 6 of 6 benchmarks
- Reasoning & science93 · #3 of 39
Based on 2 of 2 benchmarks
- Knowledge & reliability91 · #4 of 39
Based on 2 of 2 benchmarks
- Vision93 · #3 of 31
Based on 2 of 2 benchmarks
Reasoning effort trade-off
The same model at each effort setting. Lower settings give up a few points for a fraction of the cost and wait.
| Setting | Intelligence Index | Cost / task | First answer |
|---|---|---|---|
| Max | 52.7 | $3.26 | 6.9 min |
| xHigh | 52.4 | $2.31 | 3.8 min |
| High | 50.9 | $1.73 | 1.6 min |
| MediumSweet spot | 49.6 | $1.54 | 5.6s |
| Low | 45.8 | $0.82 | 3.0s |
Measured by Artificial Analysis. First answer is the median wait until the first answer token, including thinking time. “Sweet spot” is the cheapest setting that keeps at least 90% of the top score for no more than 60% of its cost.
Benchmark results
All third-party results on record. Rank is among the tracked models that have been tested on that benchmark.
Overall
#4 of 38· Max effort· Artificial Analysis
- Vals Index63.1%
#5 of 26· Max effort· $18.5 per test· Vals AI
- Arena · Text1477 ±7
#12 of 33· Max effort· 9,156 votes· Arena
Coding
- Terminal-Bench 4.059.1%
#3 of 38· Max effort· Artificial Analysis
- SciCode56.5%
#13 of 38· Max effort· Artificial Analysis
- Arena · WebDev1788 ±11
#2 of 34· Max effort· 5,057 votes· Arena
- Vibe Code Bench89.6%
#6 of 26· Max effort· Vals AI
- Code Migration68.5%
#4 of 26· Max effort· Vals AI
Agents & knowledge work
- GDPval-AA52.1
#17 of 39· Max effort· Artificial Analysis
- Arena · Agent+12.3 ±2.2
#4 of 29· Max effort· Arena
- Finance Agent v253.5%
#14 of 26· Max effort· Vals AI
#6 of 26· Max effort· Vals AI
- Legal Research Bench39.4%
#17 of 26· Max effort· Vals AI
- Tax Agent Bench63.3%
#14 of 26· Max effort· Vals AI
Reasoning & science
- Humanity's Last Exam54.7%
#5 of 39· Max effort· Artificial Analysis
- CritPt31.7%
#1 of 39· Max effort· Artificial Analysis
- ARC-AGI-2saturating95.0%
#1 of 16· Max effort· $1.12 per task· ARC Prize
- ARC-AGI-362.7%
#1 of 6· Max effort· ARC Prize
- GPQA Diamondsaturating96.1%
#1 of 27· Max effort· Artificial Analysis
Knowledge & reliability
#2 of 39· Max effort· Artificial Analysis
- AA-LCR80.7%
#19 of 39· Max effort· Artificial Analysis
Vision
- MMMU-Pro86.9%
#2 of 27· Max effort· Artificial Analysis
- Arena · Vision1284 ±11
#6 of 19· Max effort· 3,175 votes· Arena
Reported by OpenAI
Numbers the lab published at launch, run on its own harness and settings. Useful context, but not independently verified and never used in the rankings on this site.
- Terminal-Bench 4.0 · best effort57.9%
- DeepSWE v1.1 · best effort74.1%
- OSWorld 2.0 · offline set, partial score72.6%
- GPQA Diamond · best effort96%
- Humanity's Last Exam · with tools57.2%
Specifications
- Lab
- OpenAI
- Released
- Sep 3, 2026
- Availability
- Generally available
- Context window
- 1,050,000 tokens
- Max output
- 128,000 tokens
- Accepts
- Text, Images, PDFs
- Knowledge cutoff
- Apr 2026
- Weights
- Closed
- Reasoning control
- Effort low to max · pro mode
- API model ID
gpt-6-astra
Pricing
- Input / 1M tokens
- $10
- Cached input / 1M
- $1
- Output / 1M tokens
- $50
- Blended (3:1)
- $20
- Measured cost / task
- $3.26
Prompts over 272K input tokens are billed at $20 in / $75 out for the whole request. Batch and Flex are half price.
Worth knowing
- OpenAI's 99.9% ARC-AGI-3 figure used its own adapter harness. ARC Prize's standard harness gives 62.7%, which is the figure used on this site.
Official sources
Closest alternatives
The models nearest to this one on overall score.
23 benchmarks tracked. See methodology and sources.