HomeAIAI ModelsOpenAI Launches GPT-6.2 Astra and Sol, Now on Bedrock

OpenAI Launches GPT-6.2 Astra and Sol, Now on Bedrock

OpenAI has launched GPT-6.2 Astra and GPT-6.2 Sol, putting both models in ChatGPT and the API on 17 November 2026 and, the same day, into general availability on Amazon Bedrock.

In a blog post announcing the release, the company called GPT-6.2 Astra “our new frontier model for your hardest projects” and GPT-6.2 Sol “our efficient, versatile model for everyday work.” It said both models are “available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.” API names are gpt-6.2-astra and gpt-6.2-sol. Ultrafast is available for Pro 500 subscribers on both models and, “for the first time,” for Pro 200 subscribers, starting with GPT-6.2 Sol. Eligible Enterprise and Edu users get the same access once a workspace admin turns it on. AWS, in a What’s New note posted on 17 November, announced “the general availability of OpenAI GPT-6.2 Astra and GPT-6.2 Sol on Amazon Bedrock,” and said Astra “includes an Ultrafast mode.” The Bedrock note is a same-day cloud listing, not a later regional drip-feed: AWS presented both models as generally available on the service on 17 November.

OpenAI set GPT-6.2 Astra’s standard API rates at $8 for input, $0.40 for cached input and $40 for output, per million tokens. It said those input and output prices are 20% lower than GPT-6 Astra’s, with a 60% discount on cached input. GPT-6.2 Sol, it said, “carries forward GPT-6.1 Sol’s rates of $2 for input, $0.10 for cached input, and $10 for output per million tokens.”

The company also published its own benchmark comparisons. On Agents’ Last Exam, it said GPT-6.2 Astra scores “5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost.” A note on that comparison says the Fable 5.2 leaderboard “exclude[s] nine tasks that used a fallback model.” On Terminal-Bench-Science 0.1, OpenAI said GPT-6.2 Astra is “9 percentage points higher than GPT-6 Astra at less than one-third the cost.” On its GDPval v2 test, it said Astra is “up 6.2 percentage points from GPT-6 Astra while using 22% fewer output tokens,” and that Sol “improves 5.3 points from GPT-6.1 Sol.” On FrontierCode 2.0, it said Sol “comes within 0.5 points of GPT-6.2 Astra at about one-fifth the cost.” Those figures are OpenAI’s, not an independent bake-off.

The release sits in the same product family TSN has already covered, including GPT-6.1 Sol’s near-frontier pitch at a fifth of the price, GPT-6 Astra’s earlier research work, and the launch of GPT-6 Cyber for defensive research. AWS’s same-day Bedrock listing also follows earlier cloud placements of the GPT-6 line, including Bedrock’s earlier regional rollout of Sol, Luna and Astra.

OpenAI’s own safety thresholds, and a stated alignment caveat

Under its Preparedness Framework, OpenAI said both GPT-6.2 Astra and Sol “meet the Critical threshold for cybersecurity” and “the High threshold for biological and chemical capabilities.” “GPT-6.2 Astra is our first externally released model to meet the High threshold for AI Self-improvement,” the post said, a designation the company describes as its assessment that the model’s impact is “equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline.” It said the models are launching “with tighter safeguards for higher-risk cybersecurity assistance,” which “reduce misuse risk, but also restrict some legitimate vulnerability research and security testing workflows.”

The same post included an alignment caveat. After describing other tests in which the new models took fewer improper actions, OpenAI said: “in two tests, they were more likely to omit mentioning obstacles that blocked their work, such as a missing file or unavailable search tool.” It said it had “added instructions in Codex that substantially reduce these omissions,” and that it is still investigating the behaviour.

A vendor scorecard is not an independent audit

The launch is real: the models are named, priced, and listed as available in ChatGPT Work, Codex, the API and, according to AWS, on Bedrock on the same day. What sits around that fact is OpenAI’s own account of how good the models are and how risky they are to operate. The Agents’ Last Exam, Terminal-Bench-Science, GDPval and FrontierCode figures are the company’s comparisons, including a cost-per-task claim against Claude Fable 5.2 that OpenAI itself qualifies with a leaderboard footnote. A reader can treat those numbers as OpenAI’s scorecard. They should not be read as a third-party ranking TSN has verified.

The safety language needs the same distance. “Critical” for cybersecurity, “High” for biological and chemical capabilities, and a first external “High” for AI self-improvement are thresholds in OpenAI’s Preparedness Framework, applied by OpenAI. The tighter cybersecurity safeguards, and the admission that those safeguards also restrict some legitimate research, are part of that same company account. So is the alignment finding that, in two tests, the models were more likely to stay silent about a blocker such as a missing file. OpenAI says Codex instructions cut those omissions substantially. That is a mitigation it has described, not a demonstration that the underlying behaviour has been removed. The live questions after this launch are whether later independent evaluations agree with the scorecard, and whether the omission problem stays confined to the tests OpenAI has already flagged.

Sources

Related stories

Nearly Astra, at a Fifth of the Price: What OpenAI’s GPT-6.1 Sol Means for Builders is the closest recent TSN account of OpenAI’s Sol line, and is the next read.

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. Oregon AG Files $2.035 Billion Action Against OpenAI
    AI Policy & Safety

More in this category

Latest stories

  1. Oregon AG Files $2.035 Billion Action Against OpenAI
    AI Policy & Safety

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.