SIAgentNews.com
← All news
DevelopingOpenAIResearch

OpenAI's GPT-6.2 Astra scores high on intelligence but carries a premium price

The new reasoning model outperforms its peers in intelligence benchmarks while maintaining a high cost per task for enterprise users.

· First reported by artificialanalysis.ai

62Intelligence Index score for GPT-6.2 Astra (Max)

The short version

  • GPT-6.2 Astra (Max) scored 62 on the Artificial Analysis Intelligence Index, beating the median score of 43 for its class.
  • The model is a reasoning model that uses chain-of-thought processing to solve complex problems before answering.
  • Input costs are $8.00 per 1M tokens and output costs are $40.00 per 1M tokens, which are both above the median.
  • The model supports a 1M token context window and can process both text and image inputs.
  • It features five different effort levels, with scores ranging from 46 at the low level to 62 at the max level.

Why it matters

For developers and businesses, GPT-6.2 Astra represents a high-performance option for complex reasoning tasks like scientific research or software engineering. However, the high cost per task means users must weigh the model's superior intelligence against its significant price premium.

Benchmark performance results

OpenAI recently released GPT-6.2 Astra, a proprietary reasoning model designed to handle complex tasks through extended thinking. According to artificialanalysis.ai, the Max version of this model achieved a score of 62 on the Artificial Analysis Intelligence Index. This index is a composite benchmark that evaluates models across several categories, including knowledge work, coding, science, and general knowledge.

This score places GPT-6.2 Astra well above the average for models in its similar price tier. While the median score for comparable models is 43, Astra's performance suggests a significant leap in capabilities. The evaluation includes specific tests such as AA-Briefcase v1.1, HLE-Diamond, and AutomationBench-AA to measure different types of agentic work.

The model's performance is not uniform across all settings. It offers five different effort levels, ranging from low to max. At the low effort level, the model scores 46 on the index, while the xhigh and max levels both maintain a score of 62. This allows users to potentially scale the model's reasoning depth based on the complexity of the specific task at hand.

Understanding the reasoning process

GPT-6.2 Astra is classified as a reasoning model, which means it does not simply predict the next word in a sequence. Instead, it utilizes chain-of-thought reasoning to work through complicated problems internally before it provides a final answer to the user. This process is particularly useful for tasks that require multi-step logic or deep analysis of technical data.

The model is multimodal, meaning it can process different types of data beyond just plain text. Specifically, it can take both text and image inputs, allowing it to analyze, describe, and answer questions about visual information. However, it remains limited to generating text as its only output format. This combination of multimodal input and deep reasoning makes it suitable for diverse professional applications.

In terms of memory, the model features a 1M tokens context window. This context window determines the amount of text or conversation history the model can keep in its active memory during a single request. A large window allows for the processing of long documents or extensive histories, which is a key requirement for complex agentic workflows in fields like software engineering or legal work.

Cost analysis and pricing tiers

While the intelligence of GPT-6.2 Astra is high, its cost structure is notably expensive compared to other models in its class. The model costs $8.00 per 1M input tokens and $40.00 per 1M output tokens based on OpenAI's API. These figures are significantly higher than the median prices of $2.00 for inputs and $10.00 for outputs for models in a similar price range.

To provide a clearer picture for developers, the average cost to evaluate the model on the Intelligence Index is $4.00 per task. When using a blended rate that accounts for a 7:2:1 ratio of cache hits, inputs, and outputs, the cost comes to $5.88 per 1M tokens. These prices reflect a premium positioning, where users pay more for the model's ability to perform complex reasoning and high-quality output.

The pricing data highlights a trade-off between performance and budget. Because the model is proprietary, OpenAI has not disclosed the specific parameter count or model size. This lack of transparency means users must rely on these performance and cost metrics to determine if the model's intelligence justifies the higher expenditure compared to smaller or more affordable open weights models.

Impact on professional workflows

The capabilities of GPT-6.2 Astra are specifically tuned for high-level professional tasks. The Intelligence Index includes evaluations for agentic knowledge work and scientific research workflows in a terminal environment. This suggests the model is being positioned for users who need to automate complex, multi-step processes rather than simple chat interactions.

For software engineers, the model can be used for agentic coding in open-source repositories. It also handles quantitative analysis on spreadsheets and documents, as well as professional document reasoning. These applications require the model to maintain a high level of accuracy and follow complex instructions without losing track of the primary goal.

The model's ability to handle visual reasoning also opens doors for industries that rely on image analysis. Because it can describe and answer questions about images, it can be integrated into workflows that require understanding visual context. However, because it is a proprietary model, organizations must consider the implications of using a closed system for these sensitive professional tasks.

Future outlook and comparisons

As the AI landscape evolves, the distinction between reasoning models and standard models will likely become more pronounced. GPT-6.2 Astra represents a move toward models that prioritize

Sources and links

First reported by artificialanalysis.ai. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news