SIAgentNews.com
← All news

OpenAI's GPT-6.2 Astra ranks sixth on new Agent Arena Index

The new model shows significant improvements in coding and cost efficiency but trails behind competitors in conversational ability.

· First reported by Crypto Briefing

Risograph poster: picture for the story 'OpenAI's GPT-6.2 Astra ranks sixth on new Agent Arena Index'
AI-made risograph poster for this story. No real people or products are shown.
$2.94median task cost for GPT-6.2 Astra

The short version

  • GPT-6.2 Astra scored 1,291 on the Agent Arena Index, beating its predecessor by 9 points.
  • The new model reduced median task costs by approximately 40% compared to GPT-6 Astra.
  • The model excels in coding and scientific benchmarks but shows a decline in chat performance.
  • OpenAI ranks third overall in the first version of the index, trailing Anthropic and Google DeepMind.

Why it matters

For businesses, the 40% cost reduction makes high-level AI agents more viable for large-scale professional work. However, the drop in chat quality suggests the model is becoming more specialized for technical tasks rather than general-purpose conversation.

The first report card

OpenAI recently released its latest flagship model, GPT-6.2 Astra, aimed at complex professional and scientific applications. The model arrived on November 17, 2026, alongside a more efficient sibling called GPT-6.2 Sol. Both models are designed to power core products like ChatGPT Work, Codex, and the OpenAI API.

The first public evaluation of the model came from the Agent Arena Index, which was launched by Arena.ai just one day after the release. GPT-6.2 Astra secured a score of 1,291, placing it in sixth position on the leaderboard. This ranking provides a snapshot of how the model performs in real-world scenarios compared to its rivals.

While the score is respectable, it did not place OpenAI at the top of the class. The company currently ranks third overall in this inaugural version of the index. It sits behind major competitors from Anthropic and Google DeepMind in terms of overall performance metrics.

Technical gains and trade-offs

The data reveals a distinct shift in the model's capabilities compared to its predecessor, GPT-6 Astra. Specifically, the new model shows noticeable improvements in coding tasks and scientific benchmarks. OpenAI reported that the model set new highs on tests such as Terminal-Bench-Science 0.1 and Agents’ Last Exam.

However, these technical gains come with a trade-off in conversational fluidity. The index shows that GPT-6.2 Astra actually performed worse at making conversation than the previous version. This suggests that as the model becomes more capable of handling complex logic, its ability to engage in natural chat may have slightly regressed.

The model's performance is measured across four specific signals to ensure a comprehensive view. These include task completion, user satisfaction, steerability, and bash reliability. Bash reliability specifically measures how well an agent can operate a command-line shell without breaking files or programs.

Dramatic cost reductions

One of the most significant updates for users is the change in pricing and efficiency. OpenAI has positioned GPT-6.2 Astra as a more affordable option for high-level work. Input and output tokens for this model cost 20% less than they did for the previous GPT-6 Astra version.

The savings become even more pronounced for repeated tasks using cached input. Cached input refers to content the model has already processed, such as a long set of instructions sent with every request. For these specific types of inputs, OpenAI is offering a 60% discount.

The Agent Arena Index highlights these savings through median task costs. The median cost for GPT-6.2 Astra came in at $2.94, which is a significant drop from the $4.91 cost of the previous model. This represents an approximate 40% lower median cost per task for users.

How the index measures success

The Agent Arena Index is designed to move beyond static benchmarks by aggregating data from over 2.9 million real user sessions. This allows researchers to see how models behave when people actually put them to work in practical environments. It focuses on how the AI handles the nuances of human interaction and technical execution.

User satisfaction is a core component of the index, comparing how often users praise a model against how often they complain about it. This provides a qualitative look at the user experience that raw scores might miss. It helps determine if the model is actually helpful or just technically proficient.

Steerability is another key metric, measuring how well the model follows directions when a user tries to redirect its behavior. For professional tools, the ability to stay on track is vital. The index combines these human-centric signals with technical reliability to create a holistic ranking of current AI agents.

The competitive landscape

The debut of GPT-6.2 Astra highlights a crowded and competitive field for AI supremacy. Anthropic currently holds the top spot on the index with its Claude Fable 5.2 model. Several other models from Google DeepMind also managed to finish ahead of OpenAI's newest release in the initial rankings.

For OpenAI, the focus appears to be on refining the model for specific professional niches. By lowering costs and improving coding reliability, the company is targeting developers and scientists who need high-performance tools. These users may prioritize the ability to run complex scripts over the ability to have a casual conversation.

Moving forward, the industry will watch how these models evolve as more data flows into the Agent Arena Index. The fact that OpenAI ranks third suggests there is still significant room for improvement in both chat quality and overall task completion. The next phase will likely involve balancing these technical strengths with the conversational polish that users expect from a flagship product.

What to watch

  • How GPT-6.2 Astra's chat performance improves in future updates.
  • Whether the 40% cost reduction leads to a surge in enterprise API adoption.
  • The next ranking cycle of the Agent Arena Index to see if OpenAI closes the gap with Anthropic.

Try it

Check the Agent Arena Index to see how your favorite models compare in 'bash reliability' and 'steerability' metrics.

Sources and links

Read the original at Crypto Briefing

First reported by Crypto Briefing. 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