Which date is your OpenAI bill planning around, the 21st or the 30th? If the tool that answers your customer emails runs on GPT-5.6 Sol, don’t budget December at today’s discounted price. OpenAI’s pricing page now says the discount ends November 30.<\/p>\n\n\n\n
The only write-up of that date I found, RuntimeWire’s on November 3<\/a>, took it from a customer email and called the email its only source. It also said OpenAI’s public pages promised the price “at least through November 21.” That was true in October. It isn’t what the page says this morning.<\/p>\n\n\n\n I read the live pricing page<\/a> on November 5, plus two copies the Internet Archive saved, from October 5<\/a> and October 26<\/a>. (The Archive is a site that keeps old versions of web pages.) Both old copies carry this sentence:<\/p>\n\n\n\n GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.<\/p><\/blockquote>\n\n\n\n Today’s page has the same sentence in the same spot, with a different ending:<\/p>\n\n\n\n GPT-5.6 Sol’s promotional pricing is available until November 30, 2026.<\/p><\/blockquote>\n\n\n\n OpenAI’s GPT-5.6 Sol model page<\/a> carries the new line too. “At least” is gone. That phrase left room for OpenAI to keep the discount going, and the new wording doesn’t.<\/p>\n\n\n\n OpenAI’s own changelog and Amazon’s announcement still do, and both are dated August 21. The changelog entry<\/a> reads “at least through November 21, 2026,” and so does Amazon’s note<\/a> on Bedrock, Amazon’s service for renting other companies’ AI models. An old post keeps its old date, so that’s no surprise. But if you reach Sol through Bedrock, I don’t know which date Amazon will use.<\/p>\n\n\n\n I don’t know when OpenAI changed its page either. It was somewhere between October 26 and this morning. And I can’t tell whether the 30th is the last discounted day or the first full-price one. “Until” doesn’t say.<\/p>\n\n\n\n It goes up by between 25% and 50%. OpenAI’s page doesn’t print the price that comes back. The last standard price it listed, in an Archive copy from August 20<\/a>, was $5 per million tokens in and $30 out. A token is a small chunk of text, a bit under a word. RuntimeWire’s email says those rates return. Today’s discounted rates are $4 in and $20 out.<\/p>\n\n\n\n So what you send in costs 25% more, and what comes back costs 50% more. Does your tool mostly read long documents, or mostly write replies? Here’s a made-up month to see the spread: 10 million tokens in, 2 million out. Today that’s $80 (10 times $4, plus 2 times $20). After the 30th it’s $110 (10 times $5, plus 2 times $30). The more of your bill that’s answers, the closer you get to 50%.<\/p>\n\n\n\n Those are standard prices for prompts under 272,000 tokens. I didn’t check the batch, flex or fast tiers. On the same page, OpenAI lists a newer model, GPT-6.1 Sol, at $2 in and $10 out. I haven’t compared the two on real work, so I can’t say it would do your job.<\/p>\n\n\n\n Ask whoever built your tool which model it calls, and whether the call goes to OpenAI directly or through Amazon. Yesterday’s note on Google’s Gemini models<\/a> starts with the same question.<\/p>\n\n\n\nWho Still Says November 21?<\/h2>\n\n\n\n
What Does the Bill Do After the 30th?<\/h2>\n\n\n\n