The latest GPT-6.2 Astra Agent Arena results place OpenAI's model at #6 on the new Agent Arena Index, one place above GPT-6 Astra (Max). An Arena.ai post on X explains that the small overall gap hides two opposite shifts: coding scores rise 50 points while chat scores fall 54.

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

Coding is the clear gain, lifting the model from #8 to #3. Chat, which covers everyday tasks such as research, tutoring and writing, drops from #8 to #11. This article covers what Arena measured, what the numbers show, and what they cannot prove.

What the Agent Arena Index measures

Agent Arena tests models on long, multi-step work rather than one-off questions. When Arena introduced its GPT-6 Astra test, it said models could reach web search, file access and a terminal while completing a workflow. Results are judged against the average model, using a method Arena calls causal tracing.

The index itself is recent. Crypto Briefing reports that it draws on more than 2.9 million real user sessions and launched a day after OpenAI released the model on November 17, 2026. Arena described it on November 19 as its main measure of overall agent performance.

GPT-6.2 Astra Agent Arena scores by category

The post's main chart compares the two Max configurations across four indexes:

Horizontal bar chart of score differences between GPT-6.2 Astra (Max) and GPT-6 Astra (Max) on Agent Arena: Overall +9, Coding +50, Work +13, and Chat −54.
Score gaps by category: GPT-6.2 Astra leads on coding and overall, while GPT-6 Astra leads on chat, as the article describes.

Media source: x.com

  • Overall: 1,291 against 1,282 (+9)

  • Coding: 1,338 against 1,288 (+50)

  • Work: 1,280 against 1,267 (+13)

  • Chat: 1,211 against 1,265 (−54)

The public leaderboard puts Claude Fable 5.2 (Max) first at 1,448. Claude Opus 5.5, two Gemini 4 entries (one labeled preliminary) and Claude Sonnet 5.5 sit above Astra. Sonnet is only five points ahead and GPT-6 Astra nine points behind, so sixth place reflects a tight group rather than a clear lead.

Arena's coding post reports a median cost per task of $4.75, which it puts 62% below the $12.55 for GPT-6 Astra. Other sources give different figures. The leaderboard lists $3.32 for Astra, while Crypto Briefing reports $2.94 against $4.91 and calls the saving roughly 40%. Check the live leaderboard before relying on any single cost figure.

Why writing and translation scores fell

The post's second chart tracks the share of tasks users confirmed as successful. Writing and translation falls from 91.2% to 85.9%, the largest drop on the chart, and the post says creative copy prompts show the sharpest fall. Documents slips from 87.3% to 84.1%, research and tutoring from 90.8% to 89.0%, and workflow automation from 92.3% to 90.7%.

Dot-and-arrow chart comparing the percentage of tasks users confirmed as successful for GPT-6 Astra and GPT-6.2 Astra (Max) across eight categories. Writing and translation falls from 91.2% to 85.9%, shown with a red arrow; coding rises fro
Success rates by task type: the largest drop is in writing and translation, while planning, coding, and image and video improve.

Media source: x.com

Gains show up in planning (88.4% to 90.4%), coding (87.7% to 89.2%), image and video (83.3% to 86.8%) and media analysis (87.8% to 89.0%). The pattern is a trade-off: planning and coding improve while most text-heavy work slips. Documents, at 84.1%, is actually the lowest Astra score on the chart, so the weakness extends beyond writing. The data show the pattern but not its cause.

What the results do and do not prove

The success figures reflect what users judged to be finished, which is a judgment rather than a fixed answer key. Arena's author bio describes its method as community-driven evaluation, so the people who take part may not match the work you do. The overall lead is nine points, and the post publishes no confidence intervals, so it is hard to tell how much of that gap is meaningful. Every figure also uses the Max setting, so other configurations may behave differently.

What the data support is narrower than a verdict. Among Arena's users, GPT-6.2 Astra did better on coding and planning and worse on chat and writing. The results do not show that it is a weaker or stronger model overall.

Who runs Arena, and does it have a stake?

The post comes from Arena, which describes its work as community-driven evaluation. Arena also runs a business. TechCrunch reports that it reached $100 million in annualized run-rate revenue eight months after launching a commercial service, and that it has raised $250 million in total. DEV Community says the public leaderboard is free and that revenue comes from a paid product, AI Evaluations, which sells analysis to model makers.

That gives Arena a commercial reason to be trusted by the companies it ranks. It also means these numbers come from an organization that sells evaluation services. The material reviewed shows no financial tie between Arena and OpenAI, but readers should weigh the results with that in mind.

How the results compare with other tests

Arena's Code Arena WebDev results point the same direction on coding. There, GPT-6.2 Astra places #3 with 1,781 points, 18 ahead of GPT-6 Astra, and ranks first in Consumer Product.

Outside Arena, AI IQ gives the model an estimated IQ of 142, ranking it #3 among 186 public models. That figure blends many tests, and only 9 of 39 composite benchmarks have published results for the model. It is an estimate built on a different method, not a direct check on Arena's user votes.

Key takeaways

  • GPT-6.2 Astra ranks #6 on the Agent Arena Index at 1,291 points, nine ahead of GPT-6 Astra.

  • Coding rose 50 points and chat fell 54, which is why the overall rank barely moves.

  • Writing and translation shows the largest drop on Arena's success chart, from 91.2% to 85.9%.

  • Judge the model by task type, and check cost figures on the live leaderboard before choosing it.

Frequently asked questions

Is GPT-6.2 Astra better than GPT-6 Astra?

It depends on the task. Arena's data show gains in coding, planning and image and video work, and declines in chat, documents and writing. Overall, the new model is ahead by just nine points.

What rank does GPT-6.2 Astra have on Agent Arena?

It ranks #6 on the Agent Arena Index. Arena's post places it #3 on the Coding Index and #11 on chat.

Why did GPT-6.2 Astra get worse at writing?

Arena's data do not explain the cause. The data show lower user-confirmed success on writing and translation, with creative copy prompts showing the sharpest fall. Any explanation would be speculation until more information is published.

Should I use GPT-6.2 Astra for coding?

The Arena data favor it for coding, and the post reports a lower median cost per task than GPT-6 Astra. Because the results rest on user votes and a self-selected mix of tasks, test it on your own work before committing.

Sources