- GPT-6.2 Astra costs $8 per million input tokens and $40 for output, 20% less than GPT-6 Astra; cached input drops from $1 to $0.40.
- Ultrafast mode costs 3 times the Standard rate, rising to 6 times from 1 March 2027.
OpenAI released two new models on 17 November, GPT-6.2 Astra and GPT-6.2 Sol, and its developer price list shows a clear cut: the more capable Astra costs 20% less than GPT-6 Astra. The same documentation, though, contains two things that can eat those savings: new rules for caching and for Ultrafast mode. I take the price list apart and do the maths on an example.
What exactly got cheaper
GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens. GPT-6 Astra cost $10 and $50. The bigger change is in the cache: reading the repeated start of a request now costs $0.40 per million instead of $1. Writing to that cache for 30 minutes costs $10 per million instead of $12.50.
The cheaper Sol does not get cheaper: GPT-6.2 Sol costs $2 and $10, the same as GPT-6.1 Sol. According to Simon Willison, Astra is now priced the same as Claude Fable 5.2. Batch requests stay at half the Standard price, so the comparison holds there too.
The bill for one agent session
Here is a simplified example of my own: an agent takes 50 steps, each sending 100,000 tokens, of which 90,000 are a fixed opening served from the cache, and receives 2,000 tokens back.
- GPT-6 Astra: about $0.29 per step, about $15.60 per session including one cache write,
- GPT-6.2 Astra: about $0.20 per step, about $10.70 per session, roughly 30% cheaper,
- GPT-6.2 Astra with the cache not working: about $0.88 per step, about $44 for the same work.
The last line shows where the real risk is. A cache that silently stops matching turns a saving into a bill four times larger.
Trap one: the order of your instructions
The migration guide says that from GPT-6.2 onwards, instructions passed in the separate top-level field are appended after the input. To make them eligible for the cache, you have to move them to the start of the conversation as a developer message. A second change: if you edit earlier messages in the history, or switch models mid-conversation, the system drops the model’s saved reasoning. OpenAI warns that answer quality can then decline and the model may use more tokens repeating earlier work.
Trap two: speed with an expiry date
Ultrafast is the fastest way OpenAI serves requests. For the new models it costs 3 times the Standard rate, which means $24 and $120 for Astra. From 1 March 2027 that rises to 6 times, or $48 and $240. Build a product on Ultrafast today and in three and a half months you will pay twice as much for the same speed. Ultrafast also runs only with US data residency, which matters if your customers expect their data to stay in Europe.
What to check before you switch
In my view the price cut is real, but only for people who review their integration:
- move fixed instructions to the start of the request and check in your usage data how many tokens come from the cache,
- avoid editing earlier messages in long agent sessions,
- if you use Ultrafast, work out the cost from March and consider background requests, which OpenAI prices 30–40% below Standard.
Early opinions on the model itself are mixed, so test it on your own tasks before switching. I collect more model price comparisons in the Model & API pricing hub.

