GPT-6.2 Astra and Sol Released: A Big Cost Drop, and How to Choose Between the Two Models

OpenAI launches GPT-6.2 dual models, reshaping competition through cost division of labor.
Just a week after Cloud's release, OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol with a specialist/generalist split. Astra targets the hardest professional and research tasks; Sol focuses on speed and low cost. Their capability gap is small but the cost gap is large: Astra's cost per task is under a third of the previous generation, and Sol trails by only half a point on real-codebase programming tests at about one-fifth the cost.
OpenAI struck back quickly, just one week after Cloud released its new model, rolling out two models at once: GPT-6.2 Astra and GPT-6.2 Sol. In his video, Bilibili creator A Hui laid out the core information from the official announcement: the focus this time isn't benchmark scores, but the shift in cost structure and the division of labor between models for different task types.

Dual-model strategy: specialists and generalists
OpenAI has drawn a clear distinction between the two models. Astra is the heavy-lifter model, aimed at the hardest professional work and scientific research scenarios; Sol is the everyday model, focused on speed and low cost, suited to routine tasks.
A Hui uses a simple analogy to explain this division of labor: a company hires two people at once—one is expensive but can take on big projects, the other is nimble and versatile—and assigns work based on the nature of the task. This approach means users no longer have to weigh capability against cost on a single model; instead, they can switch based on task complexity.
According to the official description, Astra sets new highs on professional work and scientific research evaluations, while Sol handles high-frequency, lightweight everyday needs. The capability gap between the two isn't huge, but the cost gap is significant—and that's precisely the variable most worth watching in this release.
"Model routing" or "model division of labor" refers to offering multiple models with different capabilities and costs within the same product system, with the user or system choosing which one to call—automatically or manually—based on task complexity. This approach has precedent in the industry, for example routing requests by input length, reasoning depth, or whether tools are invoked. Its core value is avoiding "using a sledgehammer to crack a nut"—high-difficulty tasks go to large models to ensure quality, while high-frequency lightweight tasks go to small models to control cost. Implementation typically requires a unified interface protocol, shared context management, and consistent output formats; otherwise, switching models creates a fragmented experience. For users, understanding the routing logic helps determine when to manually specify a model and when to rely on the default policy, thereby balancing quality and expense.
Cost is the real headline of this release
Many people assume a new model just means higher benchmark scores, but what OpenAI repeatedly emphasizes is that it's cheaper. According to the official announcement, when Astra completes professional tasks, the cost per task is less than one-third that of the previous generation, GPT-6 Astra.

In programming tests on real codebases, Sol and Astra differ by only half a point, while Sol's cost is about one-fifth of the latter's. This comparison shows that for the vast majority of programming and everyday work tasks, Sol's value-for-money advantage is very clear.
On cybersecurity tests, Astra scores twenty-two points higher than the previous generation—a substantial capability improvement. But A Hui also cautions that these results are controlled and published by OpenAI itself, and the direct comparison coverage with Cloud gives no specific numbers, so cross-model comparisons still warrant caution.

"Cost per task" is a metric closer to actual use than "price per million tokens." A single task often involves multiple rounds of reasoning, tool calls, and retries, and token consumption varies by task, so comparing unit prices alone can be misleading. Costing by task is equivalent to converting the model's average expense in a real workflow into a single delivered result, reflecting value for money more intuitively. However, this measure also depends on the choice of evaluation task set: the more complex the task and the more retries, the higher the absolute cost; and if different vendors define tasks differently, cross-vendor comparisons lack comparability. So when you see statements like "cost reduced to one-third," you need to also consider task type, success rate, and whether retry overhead is included to judge whether that number holds in your own scenario.
How ordinary people should use it: three practical tips
For ordinary users and enterprise users, A Hui offers specific usage paths.
First, the entry points are ChatGPT's Work mode and Code mode, available to Plus, Pro, Enterprise, and Education users; the regular chat interface is not yet open.
Second, for complex projects, prioritize Astra—tasks like making PPTs, processing spreadsheets, building websites, and modifying legacy code that require deep reasoning and multi-step execution.
Third, for everyday Q&A and writing/editing, Sol is enough—faster and more economical on quota.

For users currently using ChatGPT, Cloud, or domestic models, the practical significance of this release is that under the same subscription you can allocate flexibly by task, saving the expensive model's quota for scenarios that truly need heavy lifting.
The stronger the capability, the more risk control must keep up
In the video, A Hui raises an easily overlooked issue: the stronger the cybersecurity capability, the greater the risk of malicious exploitation. OpenAI has accordingly added stricter restrictions this time.
This reflects an industry trend—competition in model capability is no longer just about benchmark scores and price; security boundaries and usage restrictions have also become part of product design. For developers and enterprise users, compliance and terms of use also need to be factored into model evaluation.
A model's score on cybersecurity evaluations typically measures the ability to find vulnerabilities, understand attack chains, or write exploit code—and such benchmarks are inherently dual-use. Improved capability means defenders can locate and fix problems faster, but it also means attackers may use the model to lower the barrier to entry. Common industry responses include: classifier interception of high-risk requests, restricting models from outputting executable attack code, introducing usage auditing and rate limiting, and providing more granular permission controls for enterprise customers. These measures sacrifice some usability to a degree, so how to trade off between security boundaries and task completion rates has become a fixed issue in model product design and a compliance dimension enterprises need to assess when procuring.
The logic of competition is changing
The competition among large models is shifting from who is smarter to who is both smarter and cheaper. OpenAI followed up quickly one week after Cloud's release, and the two companies' pace of catching up on capability and cost has clearly accelerated.
For users, this competition is a good thing: the same budget can accomplish more tasks, and model choices are closer to real workflows. As for whether OpenAI can catch up to Cloud this round, A Hui leaves the question to the audience—but what's certain is that cost efficiency has become the main battleground of the next phase of competition.
Related articles

GPT-6.2 and Astra Launch Together: OpenAI Has Finally Figured Out Its Release Cadence
OpenAI releases GPT-6.2 Sol and Astra together, with Plus users able to try them in ChatGPT Work and Codex. This article breaks down the positioning differences between the two models, Astra's practical use in long-document retrieval and rewriting, and how the simultaneous dual-model launch affects user workflows.

GPT-6.2 Dual-Model Hands-On: How Big Is the Gap Between Astra and Sol in Clock-Making and Game-Building?
GPT-6.2 Astra vs Sol hands-on: Blender clock, neon shooter, and Formula racing compared. Astra spends $419 with strong execution; Sol costs $197 but drops to 2 FPS. Both still trail Fable and Opus in games and aesthetics.

The Father of Vibe Coding's Multi-Model Strategy: Replicate His AI Coding Setup for Free
The creator of Vibe Coding never trusts a single AI model—he uses multi-model division of labor, with one writing code and one reviewing. This article analyzes his free, replicable multi-model setup and the risks of single-model tools removing models.