DEV Community

IA Compa
IA Compa

Posted on Originally published at iacompa.com

OpenAI launches GPT-6.2 Astra: frontier model for professional work at 20% lower API prices

OpenAI has launched GPT-6.2 Astra, a new flagship model intended for high-intensity knowledge and professional work. It is available in ChatGPT and through the API, and it replaces GPT-5.2 in the ChatGPT lineup.

Astra is positioned as the default model for demanding work, improving GPT-5.2 on benchmarks such as SWE-Bench Pro, GDPval and multiple GPQA variants. At the same time, it reduces API prices by around 20% versus GPT-5.2.

What changed

Astra is the successor to GPT-5.2, aimed at high-intensity professional work. According to OpenAI, the model is designed to be:

  • More capable in long-horizon agentic tasks
  • More accurate in STEM, including code and physics
  • More helpful in economics, law, health and finance
  • Less prone to hallucination, according to internal tests

Metrics

Metric GPT-6.2 Astra GPT-5.2
SWE-Bench Pro 49.3% 44.4%
GDPval (1 eval) 79.3% 73.8%
GPQA Diamond 94.8% 92.4%
GPQA Diamond (0-shot, non-thinking) 85.6% 84.1%
Metric GPT-6.2 Astra GPT-5.2
Token context 400K 400K
Input price per 1M tokens $1.40 $1.75
Cached input price per 1M tokens $0.14 $0.175
Output price per 1M tokens $11.20 $14.00

OpenAI also reports that Astra can solve GDPval tasks 1.5x faster than GPT-5.2.

Reasoning effort

  • High: GDPval 79.3%, GPQA Diamond 94.8%, SWE-Bench Pro 49.3%
  • Medium: GDPval 79.5%, GPQA Diamond 93.4%, SWE-Bench Pro 48.2%
  • Low: GDPval 77.0%, GPQA Diamond 90.3%, SWE-Bench Pro 41.9%

Benchmarks & resources

Why it matters

Astra is presented as both more capable and cheaper to use. That combination is relevant for teams that run the model at scale: the benchmarks cover professional work, coding and STEM reasoning, while the price change lowers the cost of the same workloads by about 20%.

Updates

  • November 17, 2026 — OpenAI announces GPT-6.2 Astra.
  • November 18, 2026 — This article was corrected to fix the GPT-5.2 token context: it is 400K, not 500K.

Sources

Originally published at iacompa.com

Top comments (0)