Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>
Nataliia KhandusenkoAI Eng
18 November 2026, 12:21
2026-11-18
"Virtually unusable": OpenAI releases GPT-6.2 Astra and Sol models, but receives a barrage of criticism from experienced users
On Tuesday, OpenAI unveiled two new AI models — the flagship GPT-6.2 Astra and the more affordable GPT-6.2 Sol. Despite strong internal test scores and a 20% reduction in the cost of the flagship via API, the release was met with a wave of sharp criticism from prominent developers, investors, and researchers who tested the models before launch.
On Tuesday, OpenAI unveiled two new AI models — the flagship GPT-6.2 Astra and the more affordable GPT-6.2 Sol. Despite strong internal test scores and a 20% reduction in the cost of the flagship via API, the release was met with a wave of sharp criticism from prominent developers, investors, and researchers who tested the models before launch.
Prices and specifications
GPT-6.2 Astra and GPT-6.2 Sol are proprietary. They are already available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers, as well as via API.
The flagship GPT-6.2 Astra is designed for complex codebases, web applications, and enterprise templates. At the standard rate, 1 million input tokens cost $8 (which is 20% cheaper than the previous GPT-6 Astra), and 1 million output tokens will cost $40.
Instead, GPT-6.2 Sol is focused on routine programming, writing texts, and multi-step workflows. Its cost remained at the same level as its predecessor 6.1 Sol — $2 for 1 million input tokens and $10 for 1 million output tokens. At the same time, both new products received the same context window of 1.05 million tokens with the ability to generate up to 128,000 tokens per output.
For requests over 272,000 input tokens, OpenAI doubles the cost of input data and caching, and multiplies the cost of output tokens by 1.5 (for Astra, the price increases to $16 and $60, respectively). The company also offers a high-speed Ultrafast tariff (three times more expensive than the standard one, and from March 1, 2027, six times more).
What do benchmarks show?
In OpenAI's own benchmarks, the GPT-6.2 Astra model outperformed the previous GPT-6 Astra in all nine categories, while the GPT-6.2 Sol outperformed the 6.1 Sol in eight out of nine (second only to the Vals Science Index). In the independent Artificial Analysis Intelligence Index, the new Astra scored 62 points (+5 points over its predecessor) and took second place overall behind the Claude Fable 5.2.
However, independent testing of the agent capabilities in Codex showed that Astra’s progress had effectively stalled, and Sol even regressed (notably dropping 14 points on the SWE-Atlas-QnA-AA test due to a tendency to miss important facts about the codebase). In addition, OpenAI’s system safety map notes that while the new models are less likely to engage in unacceptable behavior in office tasks, they are more likely to hide from the user the obstacles that blocked their work.
What are the new models criticized for?
The main complaints from experienced users concern the models' poor controllability, their tendency to stop prematurely on difficult tasks, and their formulaic nature in creative work, VentureBeat reports .
Matt Schumer (investor and former CEO of HyperWrite AI) said he wouldn’t even write the usual detailed review because the new models “were virtually unusable.” According to him, when he assigned Astra to work as a subagent, Claude’s models had to completely rework her work.
Theo Brown (a developer and YouTube blogger) called Astra 6.2 "the worst model OpenAI has ever released." During testing, he tried to rewrite an experimental TypeScript compiler in Rust: after seven days of work, Astra did not make any improvements, and its attempt to reduce the code size slowed down the program by 56%. Brown also noticed a strong pattern - when he and another tester independently asked Astra to create levels for different games (Outer Wilds and Majora's Mask), both received water levels with music and bells under the same name "Bellwether". At the same time, Brown noted that Astra is good at finding bugs during code reviews, is strong in 3D, and helped him design a reverse proxy server that will save him $ 25,000 per month.
Ethan Mollick (a professor at the Wharton School of Business) called this release OpenAI’s first flagship since GPT-4.5, which doesn’t seem like an obvious step forward. Despite the fact that Astra was able to create a playable space elevator with a weightless simulation in Blender and Unity, the professor described the model as “very uneven and weird” and significantly harder to control than competitors.
Dan Schipper (CEO of Every) praised the basic design sense of the Astra, but pointed out its excessive laconicism, which causes the model to miss important explanations in the texts, as well as an intrusive "corporate style" that is difficult to adapt to different brands.
At the same time, some experts came to the defense of the release: Raindrop.ai CTO Ben Hylak and AI advisor Ash Magalhaes praised GPT-6.2 Astra for working with animation, motion graphics, visual effects (VFX), and agent-based video editing. OpenAI itself added that many of their employees already use Sol as their main working tool.