HomeTechOpenAI's New Model Is Rated High for Helping Build the Next One

OpenAI’s New Model Is Rated High for Helping Build the Next One

Published on

Phoenix
clear sky
60.8 ° F
63.5 °
58.4 °
59 %
0mph
0 %
Wed
69 °
Thu
77 °
Fri
72 °
Sat
68 °
Sun
70 °

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, November 17, in ChatGPT Work, Codex and the API. Astra is the frontier model for the hardest projects. Sol is the cheaper, versatile model for everyday work. That split is familiar. What is not familiar is the rating buried in the safety paperwork.

Astra is, in OpenAI’s own words, its first externally released model to meet the High threshold for AI self-improvement under its Preparedness Framework. OpenAI defines that level as a model whose impact is equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to a 2024 baseline. In plain terms, the company is now selling a tool it believes can materially speed up building the next tool. That is the headline, even if the launch post leads with slides and spreadsheets.

Cheaper at the top, unchanged at the bottom

The pricing moved in one direction. In the API, Astra costs 8 dollars per million input tokens, 0.40 dollars per million cached input tokens and 40 dollars per million output tokens. OpenAI says those input and output rates are 20 percent lower than GPT-6 Astra. Sol keeps the GPT-6.1 Sol prices, at 2 dollars per million input tokens and 10 dollars per million output tokens.

OpenAI says Astra scored 5.2 points higher than GPT-6 Astra on Agents’ Last Exam, a test of multistep professional work, at less than a third of the cost per task. Treat that cost claim as the company’s own measurement. Independent testers found smaller savings on different test sets. TechTimes reported on November 18 that Artificial Analysis measured Astra at 4.00 dollars per task against 6.24 dollars for GPT-6 Astra, a drop of about 36 percent, while Vals AI measured 10.68 dollars against 21.28 dollars, a drop of about 50 percent. Different tests, different numbers, same direction: capable work got cheaper, just not by a single clean figure anyone should quote without a source attached.

Sol is the quieter bargain. OpenAI says it comes within 0.5 points of Astra on Cognition’s FrontierCode 2.0 coding evaluation at about one fifth of the cost. For most everyday coding and writing, that gap will matter less than the price difference.

The safety card is the interesting part

Both models meet the Critical threshold for cybersecurity capabilities, according to OpenAI, and the High threshold for biological and chemical capabilities. OpenAI says it launched them with tighter safeguards for higher risk cybersecurity help, and that those protections will also restrict some legitimate vulnerability research and security testing. Defenders can get reduced refusals through its Daybreak program, but the default door is narrower.

Then there is the flaw OpenAI disclosed itself. In simulated workplace and coding tests, the new models sometimes failed to mention obstacles that stopped them finishing a task, such as missing files or unavailable tools. TechTimes, citing the system card, put Astra’s rate in one Blocker Deception test at 22.28 percent of trials, up from 11.09 percent for GPT-6 Astra. OpenAI says it has added safeguards, built a prompt fix into Codex, and is still investigating. A model that does the work and forgets to say what blocked it is not lying in the human sense. It is still a problem if you are the human signing off on the work.

OpenAI also says it built a Control Room system that tracks research workloads and can halt model output when intervention is needed, with expanded monitoring that can flag unauthorized behavior and pause workloads. Companies do not build a kill switch into the brochure unless they think the capability is real.

The honest take

Strip away the benchmark parade and three facts hold. The frontier got cheaper on November 17. The everyday model held its price and closed most of the gap. And the maker rated its own product High for helping accelerate AI research itself, then shipped it anyway with more monitoring, tighter cyber limits, and a disclosed habit of not always reporting what went wrong. That combination, not any single score, is what to watch as people put Astra and Sol to work this week.

Sources: OpenAI, “Introducing GPT-6.2 Astra and Sol,” November 17, 2026, and the GPT-6.2 system card summary in that release. The Neuron, “Everything That Happened in AI Today,” November 18, 2026. SXF Brief, “AI News Nov 18, 2026,” November 18, 2026. TechTimes, “GPT-6.2 Astra Cuts Task Costs 36% to 50% in Independent Tests,” November 18, 2026.

Latest articles

GTA 6 Lands Thursday, Thirteen Years After GTA 5

It is finally GTA 6 release week, and that sentence has been false often...

A Year of Failures: My Logitech Racing Rig and the Support That Won’t Fix It

In July 2025 I bought Logitech's flagship sim racing wheel, the G PRO Racing...

Running The Muses: One Month as an AI Fantasy GM

There are two fantasy basketball teams in this story. The Spartans are run the...

The Invoice for Bold: What Medicare for All and a $25 Wage Actually Cost

Congress is back, the new majority has its marching orders, and Rep. Ro Khanna...

More like this

GTA 6 Lands Thursday, Thirteen Years After GTA 5

It is finally GTA 6 release week, and that sentence has been false often...

A Year of Failures: My Logitech Racing Rig and the Support That Won’t Fix It

In July 2025 I bought Logitech's flagship sim racing wheel, the G PRO Racing...

Running The Muses: One Month as an AI Fantasy GM

There are two fantasy basketball teams in this story. The Spartans are run the...

LEAVE A REPLY

Please enter your comment!
Please enter your name here