Friday, November 20, 2026

Top 5 This Week

Related Posts

GPT-6.2 Astra and Sol: OpenAI lowers its prices and follows Claude Fable 5.2, with limits


The race for AI models leaves no room for respite. A week after the release of Claude Fable 5.2 by Anthropic, OpenAI launched the November 17, 2026 two new models: GPT-6.2 Astraintended for the most demanding professional work, and GPT-6.2 Grounddesigned for daily use. OpenAI’s message comes down to one idea: best-in-class performance at a significantly lower cost. The first independent evaluations partly confirm this promise, but also point to weaknesses.

Two models, two uses

Astra targets complex tasks in science, law, finance or design. OpenAI says it has trained it to respect the conventions of each profession: its presentations use the layout, colors and typography of a provided template, down to the exact size of the fonts, from a simple screenshot. Slide titles are supposed to summarize the main idea, text to focus on the main point. OpenAI even presents it as “the best model in the world for scientists”.

Ground is presented as the all-purpose model, fast and versatile, for writing, programming or answering everyday questions. According to OpenAI, both models accomplish the same tasks with fewer chipswhich allows you to do more before reaching the usage limits of your subscription. In Codex, Astra would be more efficient for modifying and testing an existing project: “early testers saw cleaner code and more focused testing, with less back-and-forth review,” says OpenAI. Developers access both models through the API, under the names gpt-6.2-astra and gpt-6.2-sol.

Availability and price

Both models are offered in ChatGPT Work and in CodexOpenAI’s programming tool, for subscribers Plus, Pro, Enterprise and Edu. The mode Ultrafastwhich generates text up to 4.5 times faster while consuming more of the quota, arrives for the first time in the Pro subscription at $200 per month with Sol; the $500 Pro subscription benefits from this on both models.

Model Entry (per million tokens) Cache entry Output (per million tokens) Intended use
GPT-6.2 Astra 8 dollars $0.40 40 dollars Demanding professional work
GPT-6.2 Ground 2 dollars $0.10 10 dollars Daily use

Astra price drops by 20% compared to GPT-6 Astra, while Sol retains the rates of GPT-6.1 Sol. To compare, the details of its competitor’s prices appear in our article on Claude Fable 5.2 and Mythos 5.2.

What independent reviews say

OpenAI highlights its own results: on Agents’ Last Exam, a test of professional tasks in several stages, Astra is ahead of Claude Fable 5.2 by 5.5 points for 5% of its cost. But the measurements of external evaluators are more informative.

Evaluator Main result
Artificial Analysis Astra second in the general index with 62 points, behind Claude Fable 5.2 (65), ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59)
Epoch AI Astra tied with Claude Fable 5.2 on its ability index (174 points), but behind on real search tasks
Vals AI Astra 7th and Sol 9th ​​in the general index; Sol is the cheapest model in the top 10, at $3.18 per task versus $10.68 for Astra

According to Artificial AnalysisOpenAI models dominate the cost-performance frontier: no other model is both cheaper and better. Sol reaches the level of the older GPT-6 Astra for about 15% of its cost per task. In mathematics, Astra solved 7 of 68 open problems on the FrontierMath Erdős test, including one with no previously known solution, compared to a previous record of 3.

Weak points noted

Independent evaluations qualify the enthusiasm on several points:

  • programming stagnates: according to Artificial Analysis, Astra does not progress on its programming agent index and Sol falls by three points, because it more often omits important facts in its answers;
  • hallucinations remain common on very difficult factual questions: Astra guesses instead of recognizing her ignorance in 59% of cases, Sol in 55%, on a test designed to be particularly demanding;
  • the law disappoints: despite a doubling of scores, Vals AI ranks Astra only 11th on a legal test and 30th on another devoted to legal research;
  • the actual tasks: Epoch AI notes that Astra missed the central issue of simple research projects and was too conservative on an estimate.

Cybersecurity under close surveillance

GPT-6.2 Astra and Sol are better at find and exploit vulnerabilitieswrite safe code and fix bugs. Both reach the threshold ” critical “ in cybersecurity of the OpenAI security framework. On an internal test which consists of transforming known browser flaws into functional attacks, Astra exceeds 50% success, guardrails disabled.

As a result, OpenAI strengthens its guardrails on risky cybersecurity requests. The company acknowledges that these protections, intended to make abuse more difficult, will also limit some legitimate searches. Verified researchers can go through its reserved access program, which NATO and ANSSI have just obtained, as we explained regarding the opening of Daybreak to Europeans.

Epoch AI also notes a point that interests security specialists: Astra is able to chain more reasoning steps without writing them, which makes its operation more difficult to monitor.

Which one to choose?

  • For daily use (writing, summaries, questions, current programming): Groundfaster and much cheaper, is sufficient in the vast majority of cases;
  • for professional documents (presentations faithful to a charter, spreadsheets, reports) or scientific work: Astra brings real gain;
  • for intensive programming: independent evaluations do not show clear progress, it is better to compare on your own projects with Claude Fable 5.2;
  • for legal research: caution, the results are disappointing and any response must be verified.

The battle is now on price

This release confirms an underlying trend: the best models from OpenAI, Anthropic and Google are now within a few points on the tests, and the difference is made on the cost per task and reliability. By lowering the price of Astra and bringing Sol closer to peak performance, OpenAI seeks to appeal to companies deploying AI at scale, for whom every task counts. For users, this competition is good news. It also requires us to keep a critical eye on the advertisements: the tests chosen by each publisher always highlight their own models, and only independent evaluations make it possible to measure the real differences.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles