OpenAI cuts prices again: should firms buy AI by results or by tokens?
OpenAI's GPT-6.2 Astra and Sol split tasks by complexity rather than novelty, so the real question for firms is which work deserves the expensive model.

OpenAI has released two models in one day, and the cheaper one is not simply the older one. The split is by task complexity, which means the same job can now be routed to different price tiers — and the bill structure changes with it.
What OpenAI actually announced
According to IT之家 on 18 November, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol. Astra targets complex professional projects; Sol targets everyday work. Standard API pricing was published alongside: Astra costs $8 per million input tokens and $40 per million output tokens, 20% lower than the previous GPT-6 Astra on both input and output, with cached input down 60%. Sol keeps the GPT-6.1 Sol pricing of $2 input and $10 output.
For companies already using AI for content, customer service and document work, the significance is not that another model has appeared. It is that one task can now be split across price tiers by difficulty, and the shape of the bill will change.
Buy by results, or buy by tokens?
One point is easy to miss: the line OpenAI drew between the two models is task complexity, not new versus old. Astra is aimed at complex professional projects, multi-step agent tasks, coding and visual design. Sol is aimed at daily work, with the emphasis on finishing tasks using fewer tokens. According to IT之家, on Cognition's FrontierCode 2.0 benchmark Sol trails Astra by less than 0.5 points, while costing about one-fifth as much per task.
What does that mean? If your everyday output — a product description, caption text for a campaign poster, a first draft of meeting minutes — does not need the reasoning depth of a flagship model, handing that work to the economy tier saves real money.
The reverse also needs stating: a low unit price is not zero cost. The per-token price is only part of the bill; what you actually spend depends on volume. A million small scattered requests can cost more than a thousand large ones. Promotional pricing can also change, and binding an entire system to one provider's temporary price is a business risk.
So the real question is not "who is cheaper" but "is this piece of work worth an expensive model". The answer differs by business, and each has to do its own arithmetic.
Prices fell. Why are many owners still not saving?
Because the saving is not in the unit price. It is in who does the work, and how much of it.
Take a hypothetical case: a local retailer that every day needs social-media posts, replies to direct messages, and a tidy-up of the questions customers have asked. Hand all of it to a flagship model and the bill still climbs with volume, however far the unit price falls. Do all of it by hand and labour and time become the other bill. The middle road — give repetitive, formatted work to tools, keep judgement and real experience with people — is where a price war actually turns into profit.
Shu Yi, executive director of Sheyu Jiye, has a view on this: AI writes the first draft, and the owner's judgement and real experience are the most valuable part of the content. At this moment that is worth more thought than the model release itself. Falling model prices cut the cost of the first draft, not the value of judgement.
To turn price cuts into profit, see which part you can catch
Once the price war is over, the real dividing line is not who has adopted the new model but who can make "steady output" run. Shu Yi makes a second point: customer acquisition through content is not about one viral post, but about whether you can produce steadily every week and be seen continuously. That names the real bottleneck for most merchants — not that they cannot use AI, but that after using it they cannot keep it going: one post today, nothing next week.
This is the problem tools such as SHEYU AGENT (舍予AI智能体) set out to solve. It works on the content-marketing stage: 16 industry advisers each cover a professional area, and 34 zero-threshold tools cover social-media posts, AI posters and copywriting, with one login on desktop, mobile and web working everywhere. Put plainly, it gathers the process from idea to a publishable piece of content into one account, so that steady output depends less on any one person's state of mind.
The boundaries should be clear too: it produces first drafts and raw material. Brand facts, product details and the pitfalls you have personally hit still have to be fed in by you. Publishing results are subject to the platform's receipt; where a platform restricts AI hosting, the material is left for you to post manually. It cannot replace your judgement, and it cannot guarantee a viral hit.
The preparation that matters is not chasing models
Models ship several rounds a month; chasing them is a losing game. Two actions are more practical.
First, sort the AI tasks in your business into layers — which are repetitive and formatted, which are judgements only you can make. Only then do you know which tier the money should go to.
Second, build "steady output" as a capability rather than treating it as a flash of inspiration. Lower prices are an external condition; whether you can catch them depends on whether you have a process inside that can actually run.
Do those two things, and then look back at today's price-cut news. Only then will you know whether it has anything to do with you.