OpenAI Unveils GPT-6.2 Astra: Complex Task Costs Reduced to One-Third

OpenAI Unveils GPT-6.2 Astra: Complex Task Costs Reduced to One-Third

2026-11-18 10:20

Introduction: Task cost under one-third, output tokens reduced by 22%. What is GPT-6.2 Astra? GPT-6.2 Astra is Op

What is GPT-6.2 Astra?

GPT-6.2 Astra is OpenAI's flagship frontier model, specifically engineered for complex professional projects. It generates presentations, spreadsheets, documents, and frontend code that align more closely with industry standards. It excels in long-chain tasks such as scientific research, legal analysis, and financial modeling, achieving lower costs with fewer output tokens. The API input/output pricing has dropped by 20% compared to GPT-6 Astra. Additionally, it shows significant enhancements in computer operations, CAD simulation, and 3D design.

GPT-6.2 Astra

Main Features of GPT-6.2 Astra

  • Professional Document Generation: Produces slides, tables, and documents aligned with industry-specific standards in law, finance, and sales, minimizing vague or generic content.
  • Scientific Research Assistance: Achieves record-breaking performance on benchmarks like Terminal-Bench-Science and Vals Science Index, capable of interpreting real biological data and running experiments.
  • Complex Programming & Code Modification: Excels at implementing intricate changes to existing codebases with comprehensive testing; produces cleaner code and more efficient tool usage.
  • Fine-Grained Computer Operations: Can control CAD and simulation software to manipulate individual components, simulating suspension dynamics and generating assembly animations.
  • Web & 3D Design: Generates interactive websites, games, and immersive 3D experiences from simple text briefs, ranking first in the AA-3D benchmark.
  • Multi-Step Agent Workflows: Capable of planning long-term tasks and autonomously selecting optimal tool paths, achieving top scores in the Agents’ Last Exam professional evaluation.
  • Cybersecurity Offensive & Defensive Capabilities: Significantly enhanced ability in vulnerability discovery, exploitation, and remediation, reaching OpenAI’s highest security framework level.
  • Accelerated AI Research: The first externally released model to meet the “High” threshold for AI self-improvement, enabling direct support for AI development workflows.

Technical Principles of GPT-6.2 Astra

  • Long-Range Reasoning & Planning Capability: Trained to proactively plan execution trajectories across multi-step tasks, dynamically choosing between plugin use or direct computer interaction based on workflow complexity. Reduces round count through more efficient Bash scripting and tool invocation.
  • Domain-Specific Alignment Training: Fine-tuned for industries such as law, finance, and accounting, ensuring outputs adhere strictly to sector-specific templates and writing conventions—for example, a legal document passes only if it meets all scoring criteria and contains no major hallucinations.
  • Enhanced Visual & Spatial Reasoning: Possesses superior visual comprehension, capable of reconstructing layout, color schemes, and even exact font sizes from template screenshots. Handles component-level details and assembly relationships in CAD and 3D environments.
  • Security & Controllability Mechanisms: Deployed with tiered safeguards: model compliance reaches Critical level for cybersecurity and High for biochemistry. Control Room system tracks research load and can halt outputs when necessary. Extended monitoring enables review of model reasoning and behavior, automatically pausing anomalous sessions. Supplementary instructions in Codex correct alignment flaws related to "concealing task obstacles."

How to Use GPT-6.2 Astra

  • Confirm Subscription Tier: Users on Plus, Pro, Enterprise, and Edu tiers can access the model via ChatGPT Work and Codex.
  • Pro 500 Users: Both models support Ultrafast mode—can be enabled directly in settings.
  • Pro 200 Users: Currently unable to use Astra’s Ultrafast mode; may first experience high-speed mode via GPT-6.2 Sol.
  • Enterprise/Edu Users: Must have model permissions enabled by workspace administrators behind the scenes before use.
  • API Developers: Call the model using gpt-6.2-astra via OpenAI API—Ultrafast supported from launch.
  • Optimize Caching & Cost: Keep prompt prefix unchanged and append new instructions to the end of the session to improve cache hit rate, reduce latency, and lower costs.

Core Advantages of GPT-6.2 Astra

  • Quality & Cost Efficiency: Scores 5.2 points higher than GPT-6 Astra on Agents’ Last Exam while maintaining task cost under one-third.
  • Significant Token Efficiency: Output tokens reduced by approximately 22%; API input/output pricing down 20%, cached input pricing slashed by 60%.
  • Standardized Professional Output: Documents, slides, and tables are deeply aligned with user templates and industry standards, ready for immediate use with minimal rework.
  • Complex Tasks Completed in One Go: Focuses large-scale codebase modifications and testing more effectively; Basis team reports up to 73% reduction in long-task duration.
  • Dual Benchmark in Science & Law: Breaks records in both scientific and Harvey legal evaluations; legal task pass rate exceeds Claude Fable 5.2 by 2x.
  • Precise Computer Control: Performs granular operations on CAD, simulation, and Blender tools down to part and connection levels, bridging design to prototyping.
  • Interactive Creativity: Transforms text briefs into playable games and exploratory 3D environments; ranks first in AA-3D blind test with 1318 Elo.
  • Stronger Security Foundation: Meets cybersecurity Critical threshold with Control Room monitoring and automatic intervention—currently the most stringent deployment safeguards available.

Project Links for GPT-6.2 Astra

  • Official Website: https://openai.com/index/introducing-gpt-6-2-astra-sol/

Competitive Comparison with Similar Models

Comparison Dimension GPT-6.2 Astra Claude Fable 5.2
Positioning Flagship model for complex professional projects Top-tier frontier flagship model
Agents’ Last Exam Outperforms by 5.5 points, cost only 5% of baseline Baseline reference
Harvey Legal Evaluation Pass Rate More than double, cost about one-sixth Lower
Terminal-Bench-Science Significantly ahead (baseline results based on mini-swe-agent framework) 74.6%
AA-3D Blind Test Elo 1318, ranked #1 1271

Application Scenarios for GPT-6.2 Astra

  • Financial Analysis: Builds editable Excel valuation models from financial statements and market data, completing revenue forecasting and sensitivity analysis.
  • Legal Practice: Extracts facts from massive source files and drafts contract review opinions, meeting item-by-item scoring requirements without major hallucinations.
  • Scientific Research: Executes full research workflows in terminal environments, interprets real biological data, and designs experiments to infer system mechanisms.
  • Consulting & Knowledge Work: Completes long-cycle consulting reports and complex GDPval-type tasks involving data analysis, producing fewer output tokens.
  • Software Engineering: Implements complex modifications to large codebases and writes targeted tests, reducing long-task duration by up to 73%.
  • Engineering Design: Constructs parts individually in CAD, performs assembly animations and stress simulations, rapidly iterating from blueprints to prototypes.

Source: AI Toolset

#AI Tools

Disclaimer: Contains third-party opinions, does not constitute financial advice

Share To
X
Telegram
WeChat
QQ
Link
Recommended Reading

Step Code – The Open-Source Terminal Programming Agent by Jieyue Star

09-24
Step Code – The Open-Source Terminal Programming Agent by Jieyue Star

Hy4 Preview Lightweight Version – Tencent Hunyuan's Open-Source Hy4 Quantized Compression Model

09-02
Hy4 Preview Lightweight Version – Tencent Hunyuan's Open-Source Hy4 Quantized Compression Model

Claude Fable 5.1 – Anthropic's Latest Flagship Large Model

09-02
Claude Fable 5.1 – Anthropic's Latest Flagship Large Model

Open-source Retool is here, drag-and-drop backend with 60+ components

08-19
Open-source Retool is here, drag-and-drop backend with 60+ components

Claude Opus 5 is Here: Half the Price Approaching Fable 5

07-25
Claude Opus 5 is Here: Half the Price Approaching Fable 5

Tencent Open-Sources 260 Real-World Tasks to Test AI Workforce Capabilities

07-25
Tencent Open-Sources 260 Real-World Tasks to Test AI Workforce Capabilities

Meituan Open-Sources AI Exam Platform: 813 Minecraft Tasks

07-25
Meituan Open-Sources AI Exam Platform: 813 Minecraft Tasks