OpenAI's GPT-6.2 Astra and Sol models face mixed early reviews
The new model updates aim to bolster OpenAI's professional standing but have faced criticism from early testers for lacking significant leaps over rivals.
The short version
- Astra ranks behind Fable 5.2 on independent benchmarks but sits ahead of Google’s Gemini 4 Carbon.
- The more affordable Sol model costs roughly 15% as much as previous versions to complete tasks.
- Astra shows no meaningful upgrade in coding, while Sol actually regressed from the GPT-6.1 version.
- OpenAI's internal assessment placed Astra at its "High" threshold for AI self-improvement capabilities.
Why it matters
This release comes at a critical time as Anthropic prepares for its IPO. OpenAI needs to maintain its lead in the professional market to avoid being overshadowed by its primary competitor.
Astra and Sol launch
OpenAI recently introduced GPT-6.2 Astra and GPT-6.2 Sol as its latest upgrades to top and everyday models. These releases were specifically designed to cater to professional workflows and address the growing competition in the AI space. The company aimed to provide a robust answer to the market dominance currently held by Anthropic's models.
Astra is positioned as a high-end model, while Sol serves as a more cost-effective alternative. According to the report from therundown.ai, Sol is priced at roughly 15% of the cost required to finish each task compared to the previous GPT-6 Astra. This pricing structure suggests a push toward making high-quality AI more accessible for high-volume professional use.
The timing of the release was significant as it occurred just days before Anthropic's public listing. OpenAI faced the pressure of proving its models could still lead the industry. However, the initial reception from the community has been mixed, with many users looking for a more substantial leap in capabilities than what was delivered.
Benchmark performance
Independent testing by Artificial Analysis shows that Astra has climbed 5 points since September. It currently sits ahead of Google’s Gemini 4 Carbon but remains behind Fable 5.2 in the rankings. These scores provide a snapshot of how the model stacks up against other major players in the current landscape.
The model demonstrated specific strengths in several technical categories. Astra topped tests in science, legal work, robotics, and 3D builds. These results indicate that the model is well-suited for complex, specialized tasks that require high levels of accuracy and reasoning in structured environments.
Despite these successes, the model struggled in other areas. It showed no meaningful upgrade in coding capabilities compared to previous versions. Furthermore, the Sol model actually showed a regression from GPT-6.1 in coding tasks, which may be a point of concern for developers who rely on these models for software engineering.
Tester feedback and critiques
Early feedback from the testing community has been notably harsh. Some users have expressed significant frustration with the quality of the output. One tester even described the new release as "the worst model OpenAI has ever released," highlighting a gap between the company's goals and user experience.
Specific complaints centered on the readability of the answers provided by the models. Users noted that the responses were sometimes difficult to parse or lacked the clarity expected from a top-tier product. This feedback suggests that while the underlying logic might be improved, the user interface and communication style may need refinement.
The criticism is particularly stinging because of the high expectations surrounding the GPT-6 series. For a company that has consistently set the standard for AI, failing to provide a clear leap forward can hurt brand loyalty. OpenAI now faces the challenge of addressing these complaints while competing for a fresh war chest from Wall Street.
Internal capabilities and future
OpenAI provided its own assessment of the new models, noting that Astra reached its "High" threshold for AI self-improvement capabilities. This marks a first for any model the company has released publicly. Self-improvement refers to the model's ability to refine its own processes and logic over time.
While the external reviews are currently mixed, these internal metrics suggest that OpenAI is working on deeper architectural improvements. The goal is to create models that can become more efficient and accurate as they are used. This focus on self-improvement is a key part of the company's long-term strategy to maintain a competitive edge.
Moving forward, the success of GPT-6.2 will likely depend on how well OpenAI can iterate based on this feedback. The company must bridge the gap between its internal technical milestones and the practical needs of its users. If they can improve coding performance and answer clarity, they may be able to reclaim the top spot on the charts.
What to watch
- How OpenAI responds to the coding regression in the Sol model.
- Whether Anthropic's IPO results lead to more aggressive competition.
- Future updates to Astra's readability and output formatting.
Try it
Check the Artificial Analysis rankings to see how Astra compares to Fable 5.2 in your specific use case.
Sources and links
- Officialdefender’s window openai.com
- OfficialOpenAI drops GPT-6.2 Astra, Sol to mixed fanfare openai.com
- Officialrolling out blog.google
- Officialblog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gt blog.google
- Postx.com/theo/status/2122757647170703550 x.com
- Postx.com/danshipper/status/2122754729012777052 x.com
- Reportingtherundown.ai: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by therundown.ai · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
