OpenAI has released the GPT-6.2 series models Astra and Sol, with sharply split pricing: Astra at $8 per million input tokens and $40 per million output tokens, and Sol at $2 and $10. According to a roundup of industry news from Nov 13 to 20 by Reddit user ksraj1001, Sol comes close to Astra on coding benchmarks at about one-fifth the cost. The same system card rates a released model's cyber capability as "Critical" for the first time, and Astra's self-improvement capability as "High".

Tiered pricing and capability positioning

Based on the published prices, the GPT-6.2 series forms a high-low pairing: Astra targets high-capability use cases, while Sol offers comparable coding performance at a lower price. The roundup labels which figures come from the vendors.

The author of the roundup, an API-based developer, argues that three frontier-class models with input prices below roughly $2 per million tokens appearing in one week means task-based routing has moved from an optimization tactic to a basic skill.

In the same week, Alibaba released Qwen4-Max: about 2.4T total parameters in a MoE architecture, roughly 71B active, 1M context, native vision, available only via API with no open weights, priced at about $1.5/$4.5 per million tokens. The company says cost per task is about 60% lower than Qwen3.8-Max, based on its own benchmarks. A smaller open-source version is "coming soon".

Safety rating and tighter guardrails

The system card's "Critical" rating for a released model's cyber capability is the safety signal drawing attention in this launch. Reddit user cardak98 reports that Sol 6.2 refused to help with a reverse-engineering request to port a 25-year-old game to the web, while 6.1 had cooperated smoothly on the same project a month earlier. The user sees this as a user-visible case of tighter safety guardrails, with behavior shifting abruptly across versions on the same task, sparking discussion about the transparency of model guardrail policies.

Early testers gave mixed feedback on the models' autonomy.

Community questions and communication pressure

AI blogger kimmonismus publicly called out OpenAI on X, questioning the decline in the quality of its community communication. He wants the company to explain why GPT-6.2 Sol and Astra are underperforming, its fix plans, timeline and roadmap, and how it will respond to competition from Google. He said OpenAI's communication with the community used to be much better than it is now, and called for substantive updates beyond "we're fine"-style brush-offs.

What to watch

  • Whether OpenAI formally responds to the performance issues with Sol and Astra and gives a fix timeline;
  • What the system card's "Critical" cyber capability rating specifically means and how it affects developers in practice;
  • Whether Sol 6.2's tighter guardrails represent a cross-version policy shift, and whether the company can improve transparency;
  • Adoption of task-based routing and multi-model combinations as frontier models trend toward lower prices.