Understand this story
On November 18, 2026, OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol, making them available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex starting that day, and simultaneously launching them on the OpenAI API. OpenAI said Astra is aimed at highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third that of GPT-6 Astra; Sol is aimed at everyday multi-step agent workflows. On the same day, ARC Prize announced evaluation results, saying Astra achieved a 100.0% SOTA score on ARC-AGI-3, becoming the first model to earn a perfect score on that benchmark, with 95.8% on ARC-AGI-2 and 98.5% on ARC-AGI-1. Noam Brown reposted a comment saying the model can be used for just $20 per month. OpenRouter subsequently launched both models and their Pro versions. Yuchen Jin questioned that Astra showed no improvement over GPT-6 Astra in coding.
AI-generated from public reports
Latest development · Report
GPT-6.2 Astra Tops AA-Robotics v0.1 with 79 PointsDevelopments
11 confirmed developments
Artificial Analysis says GPT-6.2 Astra tops AA-Robotics v0.1
Source:Artificial Analysis- Source:Arena
Artificial Analysis benchmark shows Astra for Law 6.2 leading
Source:Artificial AnalysisNoam Brown says GPT-6.2 Astra surpasses top human intelligence in some domains
Source:Noam BrownARC Prize announces GPT-6.2 Astra achieves 100% on ARC-AGI-3
Source:ARC PrizeARC Prize announces GPT-6.2 Astra achieves 100% on ARC-AGI-3
Source:ARC PrizeARC Prize announces GPT-6.2 Astra achieves 100% on ARC-AGI-3
Source:ARC Prize- Source:ARC Prize
Yuchen Jin questions whether GPT-6.2 Astra shows no improvement in coding capability
Source:Yuchen Jin- Source:OpenRouter
- Source:OpenAI
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 18
Artificial AnalysisSelected
GPT-6.2 Astra Tops AA-Robotics v0.1 with 79 Points
In Artificial Analysis's AA-Robotics v0.1 evaluation, GPT-6.2 Astra tops the leaderboard with 79 points, 10 points higher than GPT-6 Astra, and ahead of GPT-6.2 Sol (78), Claude Fable 5.2 (75), and Gemini 4 Hydrogen (58). The test involves controlling a real robotic arm through Codex; GPT-6.2 Astra passed six of eight physical tasks, partially completed the liquid-pouring task, and failed only the box-measuring task.
Noam BrownSelected
Noam Brown shares and comments on GPT-6.2 Astra: available for $20 per month
Noam Brown reposted OpenAI's release of GPT-6.2 Astra, saying it can be used for $20 per month and that in some fields it surpasses top human intelligence; he is most looking forward to seeing people use it to make scientific discoveries. OpenAI's original post said the model is aimed at the hardest projects, sets new highs on evaluations across professional work and scientific research, and costs less than a third per task of GPT-6 Astra, with the difference showing in tasks such as slideshows, websites, and 3D design.
OpenRouterFirst-partySelected
OpenAI's GPT-6.2 Astra and GPT-6.2 Sol are live on OpenRouter
OpenAI's GPT-6.2 Astra and GPT-6.2 Sol are now live on OpenRouter, and both come in Pro versions. Astra targets the hardest projects, while Sol targets everyday agent work; according to information cited by OpenRouter, Astra is priced at $8 per million input tokens and $40 per million output tokens, while Sol is priced at $2 input and $10 output.
Artificial Analysis
OpenAI Astra for Law 6.2 Tops Harvey LAB-AA v1.1
OpenAI's Astra for Law 6.2 ranks first on the Harvey LAB-AA v1.1 legal benchmark with a 20.8% Hallucination-Gated All-Pass Rate, while GPT-6.2 Astra (max) places second at 15.0%. Astra for Law 6.2 is a specialized configuration of GPT-6.2 Astra, paired with an expanded legal retrieval index and instructions for research, weighing precedents, and drafting answers. The benchmark comprises 120 private tasks across 24 practice areas and penalizes model hallucinations in its primary metric.
Yuchen Jin
Yuchen Jin Questions Whether GPT-6.2's Coding Ability Has Improved
Yuchen Jin posted a question: GPT-6.2 Astra shows no improvement in coding compared with GPT-6 Astra, while GPT-6.2 Sol is actually worse at coding than GPT-6.1 Sol, and he went on to ask, "How is this possible?"
Greg BrockmanSelected
OpenAI releases GPT-6.2 Astra and Sol
Greg Brockman reposted OpenAI's announcement, stating that GPT-6.2 Astra and Sol are now live, emphasizing high efficiency and being aimed at real-world scenarios. He added that these two models perform excellently in professional work and scientific research, with per-task costs greatly reduced.
ArenaSelected
OpenAI GPT-6.2 Astra and Sol Launch on Arena for Evaluation
Arena announced that OpenAI's GPT-6.2 Astra and Sol have entered Arena and are available to try in Agent Arena, with scores to be announced later. Agent Arena evaluates models based on millions of real long-horizon agent tasks submitted by users worldwide; models can call web search, file system, and terminal tools, and the leaderboard uses a causal tracing method to measure each model's performance relative to the average model. These two models have also appeared in Code Arena's WebDev, Text, Vision, Search, and Document.
ARC PrizeFirst-partySelected
ARC Prize announces GPT-6.2 Astra's results on the ARC-AGI series of benchmarks
ARC Prize announced OpenAI GPT-6.2 Astra's results on the ARC-AGI series of benchmarks: ARC-AGI-3 reached 100.0% (SOTA), at a cost of $4.1K (provider adapter), and 70.3% at $24.2K in standard mode; ARC-AGI-2 scored 95.8% (SOTA) at $0.50 per task; ARC-AGI-1 scored 98.5% at $0.17 per task. ARC Prize stated that this is the first model to achieve a 100% score on ARC-AGI-3. Compared with Sol's best ARC-AGI-3 result of 98.2% under high reasoning, Astra is 1.8 percentage points higher under max reasoning, but costs 64% more. On the 25 public ARC-AGI-3 games, Astra used 64.1% fewer actions than the human baseline and 5.2% fewer than GPT-6 Astra.
ARC PrizeFirst-partySelected
ARC Prize announces GPT-6.2 Astra's results on the three ARC-AGI benchmarks
ARC Prize announced the results of OpenAI GPT-6.2 Astra on the three ARC-AGI benchmarks: ARC-AGI-3 reached 100.0% (SOTA), with a cost of $4.1K under the provider adapter and 70.3% at $24.2K under the standard setting; ARC-AGI-2 scored 95.8% (SOTA) at $0.50/task; ARC-AGI-1 scored 98.5% at $0.17/task. ARC Prize states that this is the first model to achieve a 100% score on ARC-AGI-3. On the ARC-AGI-3 semi-private games, GPT-6.2 Astra used 53% fewer first-action reasoning tokens than GPT-6 Astra; on the ARC-AGI-2 semi-private tasks at the low to xhigh levels, it used 51% fewer reasoning tokens, while at the max level it used 12% more.
ARC PrizeFirst-partySelected
ARC Prize announces evaluation results of GPT-6.2 Astra on ARC-AGI
ARC Prize announces the results of OpenAI GPT-6.2 Astra on the ARC-AGI series of benchmarks: ARC-AGI-3 reaches 100.0% (SOTA), with a cost of $4.1K using the provider adapter, and 70.3% at $24.2K with the standard approach; ARC-AGI-2 is 95.8% at $0.50/task, and ARC-AGI-1 is 98.5% at $0.17/task. ARC Prize states that this is the first model to achieve a 100% score on ARC-AGI-3. On the 25 public ARC-AGI-3 games under the max reasoning setting, GPT-6.2 Astra uses 64.1% fewer actions than the human baseline and 5.2% fewer than GPT-6 Astra; on ARC-AGI-2 semi-private tasks at the low to xhigh levels, its reasoning token usage is 51% lower than the previous generation, while at the max level it is 12% higher.
ARC PrizeFirst-partySelected
ARC Prize announces evaluation results of GPT-6.2 Astra and Sol on ARC-AGI
ARC Prize announces that OpenAI GPT-6.2 Astra achieved a SOTA score of 100.0% on ARC-AGI-3, becoming the first model to earn a perfect score on that benchmark, with 95.8% on ARC-AGI-2 and 98.5% on ARC-AGI-1. Sol reached a maximum of 98.2% on ARC-AGI-3, 1.8 percentage points lower than Astra, but at 64% lower cost under the provider adapter approach. ARC Prize also provides links to reproducible results, testing policy, and the complete results for Astra and Sol.
ARC PrizeFirst-partySelected
ARC Prize evaluation: GPT-6.2 Astra uses 64.1% fewer actions than the human baseline on ARC-AGI-3
ARC Prize published the evaluation results for GPT-6.2 Astra on ARC-AGI-3: across 25 public games under the max reasoning setting, its action count was 64.1% lower than the human baseline and 5.2% lower than GPT-6 Astra. On ARC-AGI-2 semi-private tasks, GPT-6.2 Astra used 51% fewer reasoning tokens than its predecessor at the low through xhigh levels, while using 12% more at the max level. In ARC-AGI-3 semi-private games, its first-action reasoning tokens were 53% lower than GPT-6 Astra's.
ARC PrizeFirst-partySelected
ARC Prize announces GPT-6.2 Astra's scores and costs on the ARC-AGI benchmarks
ARC Prize announces the results of OpenAI GPT-6.2 Astra on the ARC-AGI series of benchmarks: ARC-AGI-3 scored 100.0% (SOTA), at a cost of $4.1K (provider adapter), 70.3% and $24.2K (standard); ARC-AGI-2 scored 95.8% (SOTA), at $0.50/task; ARC-AGI-1 scored 98.5%, at $0.17/task. On the ARC-AGI-2 semi-private tasks, from the low to xhigh tiers, GPT-6.2 Astra's reasoning token usage is 51% lower than the previous generation GPT-6 Astra, while at the max tier it is 12% higher. On the 25 public games in ARC-AGI-3 under max reasoning, its number of actions is 64.1% lower than the human baseline and 5.2% lower than GPT-6 Astra.
ARC PrizeFirst-partySelected
ARC Prize verifies that GPT-6.2 Astra reaches 100% on ARC-AGI-3
ARC Prize has published verification results for OpenAI GPT-6.2 Astra on ARC-AGI: it scores 100.0% on ARC-AGI-3 (SOTA), 95.8% on ARC-AGI-2, and 98.5% on ARC-AGI-1. On ARC-AGI-3, the cost under the provider adapter is $4.1K, while under the standard setting the score is 70.3% at a cost of $24.2K; the per-task costs for ARC-AGI-2 and ARC-AGI-1 are $0.50 and $0.17, respectively. ARC Prize states that this is the first model to achieve a 100% score on ARC-AGI-3, and provides a comparison with the previous-generation GPT-6 Astra and Sol in terms of the number of actions and reasoning token usage.
ARC PrizeFirst-partySelected
ARC Prize publishes score and cost comparison of Sol and Astra on ARC-AGI-3
ARC Prize reports that Sol achieved a score of 98.2% on ARC-AGI-3 at the high reasoning setting, 1.8 percentage points lower than Astra's 100% at the highest reasoning setting, but at 64% lower cost under the provider adapter setup. Sol scored 85.8% on ARC-AGI-2 at $0.13 per task, and 97.5% on ARC-AGI-1 at $0.06 per task; under the ARC-AGI-3 standard setup it scored 54.9% at a cost of $5.5K.
ChatGPTFirst-partySelected
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol, which the company says are highly efficient, skilled at creating slide decks and building websites, and also capable of handling previous work tasks. Starting the same day, the two models are available to Plus, Pro, Enterprise, and Edu users and can be used in ChatGPT Work and Codex.
Testing CatalogSelected
OpenAI launches GPT-6.2 Astra and… on ChatGPT Work, Codex, and API
OpenAI is rolling out two models, GPT-6.2 Astra and GPT-6.2 Sol, on ChatGPT Work, Codex, and the API. GPT-6.2 Sol scored 57.3 on the Artificial Intelligence Index. Pro 200 users can use GPT-6.2 Sol Ultrafast, while Pro 500 users can use Ultrafast on both the GPT-6.2 Sol and Astra models.
OpenAIFirst-partySelected
OpenAI Releases Two Models: GPT-6.2 Astra and Sol
OpenAI has released two models, GPT-6.2 Astra and GPT-6.2 Sol. Astra targets highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third that of GPT-6 Astra; Sol targets everyday multi-step agentic workflows. Astra scores 22 points higher than GPT-6 Astra on ExploitBench Internal Port, so OpenAI has added stricter safety measures for high-risk cybersecurity assistance and plans to let approved defenders use these two models in Daybreak with fewer refusals. Both models are now available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and developers can call gpt-6.2-astra and gpt-6.2-sol via the OpenAI API.
OpenAIFirst-partySelected
OpenAI Releases GPT-6.2 Astra and Sol, Launching on ChatGPT Work, Codex, and API
OpenAI has released GPT-6.2 Astra and Sol, available starting today in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and developers can call gpt-6.2-astra and gpt-6.2-sol via the OpenAI API. Astra targets highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third that of GPT-6 Astra; Sol targets everyday multi-step agent workflows. Pro 500 subscribers can use Ultrafast on both models, and OpenAI is also for the first time opening Ultrafast to Pro 200 subscribers, starting with GPT-6.2 Sol. Astra scores 22 points higher than GPT-6 Astra on ExploitBench Internal Port, and OpenAI has therefore added stricter safeguards for high-risk cybersecurity assistance on both models.
OpenAIFirst-partySelected
OpenAI releases two models: GPT-6.2 Astra and GPT-6.2 Sol
OpenAI releases two models, GPT-6.2 Astra and GPT-6.2 Sol. Astra targets highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third that of GPT-6 Astra; Sol focuses on multi-step agent workflows in everyday work. Both models are available starting today in ChatGPT Work and Codex to Plus, Pro, Enterprise, and Edu users, and developers can call gpt-6.2-astra and gpt-6.2-sol via the OpenAI API, with Ultrafast support; Pro 200 subscribers can use Ultrafast for the first time, starting with Sol. Astra scores 22 points higher than GPT-6 Astra on ExploitBench Internal Port, so OpenAI has added stricter protections for high-risk cybersecurity assistance on both models.
OpenAIFirst-partySelected
OpenAI Releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol. Astra targets highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third that of GPT-6 Astra; Sol targets everyday multi-step agent workflows, and many OpenAI employees use it as their go-to model. Both models are available starting today in ChatGPT Work and Codex to Plus, Pro, Enterprise, and Edu users, and developers can call gpt-6.2-astra and gpt-6.2-sol through the OpenAI API, with Ultrafast support. Astra scores 22 points higher than GPT-6 Astra on ExploitBench Internal Port, and OpenAI has therefore tightened protections for high-risk cybersecurity uses.
OpenAIFirst-partySelected
OpenAI releases two models: GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol, now available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Astra targets highly difficult projects, setting new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third of GPT-6 Astra; Sol targets everyday multi-step agent workflows. Both models launched simultaneously on the OpenAI API; Pro 500 subscribers can use Ultrafast, and Pro 200 subscribers receive Ultrafast for the first time, starting with Sol. Astra scores 22 points higher than GPT-6 Astra on ExploitBench Internal Port, so OpenAI added stricter cybersecurity protections to both models.
OpenAIFirst-partySelected
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol. Astra targets highly difficult projects and sets new highs on professional work and scientific research evaluations, with a per-task cost of less than one-third of GPT-6 Astra; Sol targets everyday work and can handle multi-step agentic workflows. Both models are now available in ChatGPT Work and Codex to Plus, Pro, Enterprise, and Edu users, and developers can call gpt-6.2-astra and gpt-6.2-sol via the OpenAI API, with Ultrafast support.
OpenAIFirst-partySelected
OpenAI Releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol. Astra targets highly difficult projects and set new highs on professional work and scientific research evaluations, with a cost per task of less than one-third of GPT-6 Astra. Sol is positioned as a general-purpose model for everyday work and can handle multi-step agentic workflows. Both models are now available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and developers can call gpt-6.2-astra and gpt-6.2-sol via the API.
9to5MacSelected
OpenAI upgrades GPT-6.2 Astra and Sol models for ChatGPT and Codex
OpenAI has launched GPT-6.2 Astra and GPT-6.2 Sol simultaneously on ChatGPT Work and Codex, and both models can complete demanding tasks using fewer output tokens. Astra remains OpenAI's recommended model for the hardest projects; it can match layout, color schemes, and font sizes from screenshot templates, and in Codex it is better at modifying and testing code within existing projects. Sol is aimed at everyday coding and writing, and on FrontierCode 2.0 it narrows the gap with Astra to within half a point at about one-fifth the cost. Both models are available starting today to Plus, Pro, Enterprise, and Edu; the $200/month Pro plan offers Ultrafast mode for Sol for the first time, and the $500/month plan gives access to both models. Ultrafast generation is up to 4.5 times faster than standard speed but consumes more quota.
WallstreetcnSelected
OpenAI Releases GPT-6.2 Astra Model, After Previously Withdrawing the Previous Update
OpenAI has released GPT-6.2 Astra, a new version of its most advanced model Astra, nearly two months after the previous update was withdrawn. OpenAI says the model is more efficient at performing highly difficult professional tasks, and that the slides, documents, and spreadsheets it generates better align with user instructions and industry standards, reading less like AI-generated content. In a benchmark test of multi-step professional work, it achieved better performance at a cost far lower than Anthropic's most powerful Fable model.
AI Trends