cd /news/ai-agents/openais-gpt-6-2-astra-ranks-sixth-on… · home › topics › ai-agents › article
[ARTICLE · art-173851] src=cryptobriefing.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

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

OpenAI's GPT-6.2 Astra debuted in sixth place with a score of 1,291 on the Agent Arena Index, which Arena.ai launched one day after OpenAI released the model on November 17, 2026. The new model scored 9 points above its predecessor, GPT-6 Astra, with median task cost falling to $2.94 from $4.91 (about 40% lower), though it posted higher coding scores and lower chat scores; Anthropic's Claude Fable 5.2 took the top spot, and OpenAI ranks third overall behind Anthropic and Google DeepMind. Arena.ai built the index from more than 2.9 million real user sessions, tracking task completion, user satisfaction, steerability and bash reliability.

recommendation

Evaluate agent models on real user sessions across task completion, satisfaction, steerability and bash reliability, not chat alone, and weigh median cost per task: GPT-6.2 Astra cut it to $2.94 from $4.91.

more in playbook: Evals and testing for LLM apps
by read3 min views8 publishedNov 21, 2026
OpenAI’s GPT-6.2 Astra ranks sixth on the new Agent Arena Index
Image: Cryptobriefing (auto-discovered)

FoxTPNL / Wikimedia Commons (CC BY 4.0) The new model beats its predecessor on coding and cost but slips on chat, debuting behind Anthropic and Google DeepMind

OpenAI’s newest flagship model has its first public report card, and the grade is respectable but not top of the class. GPT-6.2 Astra debuted at sixth place on the Agent Arena Index with a score of 1,291.

The ranking comes from Arena.ai, which launched the index just one day after OpenAI released the model. The result is a model that got noticeably better at writing code and noticeably worse at making conversation.

What OpenAI shipped #

OpenAI introduced GPT-6.2 Astra on November 17, 2026. The company positions it as an advanced model built for complex professional and scientific work.

It arrived alongside a sibling, GPT-6.2 Sol, which is the more efficient option of the pair.

Both models are aimed at OpenAI’s main products: ChatGPT Work, Codex, and the OpenAI API.

Pricing is a big part of the pitch. Input and output tokens for GPT-6.2 Astra cost 20% less than they did for GPT-6 Astra, its predecessor.

Cached input gets a steeper cut, with a 60% discount. Cached input refers to content the model has already processed and can reuse, like a long instruction set sent with every request.

OpenAI has also said the model set new benchmark highs on tests including Agents’ Last Exam and Terminal-Bench-Science 0.1. The company reported reaching those scores at lower costs per task than its previous model managed.

AI, tech, and the markets they move—in one daily briefing.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

How the Agent Arena Index works #

The Agent Arena Index aggregates data from more than 2.9 million real user sessions, measuring how models behave when people actually put them to work.

The index tracks four signals. The first is task completion, which asks whether the agent finished the job.

The second is user satisfaction, measured by comparing how often users praise a model against how often they complain about it. The third is steerability, meaning how well the model follows direction when a user tries to redirect it.

The fourth is bash reliability. Bash is the command-line shell developers use to run programs and manage files, so this measures whether an agent can operate a terminal without breaking things.

The scorecard #

GPT-6.2 Astra’s 1,291 score puts it 9 points ahead of GPT-6 Astra overall.

The new model posted higher coding scores than its predecessor but lower chat scores.

Cost efficiency is where the upgrade looks most dramatic. Median task cost came in at $2.94 for GPT-6.2 Astra, compared to $4.91 for GPT-6 Astra.

That works out to approximately 40% lower median cost per task.

The top spot went to Anthropic’s Claude Fable 5.2. Several models from Google DeepMind also finished ahead of OpenAI’s new release.

At the company level, OpenAI ranks third overall in this first version of the index, behind Anthropic and Google DeepMind.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

── more in #ai-agents 4 stories · sorted by recency
── more on @openai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/openais-gpt-6-2-astr…] indexed:0 read:3min 2026-11-21 · —