Every vendor charging for AI add-ons is overcharging you, and OpenAI just proved it

OpenAI has released GPT-6.1 Sol, a model it says nearly matches its flagship GPT-6 Astra on coding, computer use and professional work. It costs $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens. Astra costs $10, $50 and $1. On OpenAI's own science benchmark at maximum effort, Sol averages $5.47 per task against $23.80 for Astra and $23.21 for Opus 5.5.

I think every B2B vendor still charging a premium AI add-on fee is now overcharging, and I think buyers should reopen those line items before the next renewal. The fees were set when the ingredient cost roughly five times what it costs today.

I went in expecting a straightforward price cut. Sol's uncached rates are unchanged from GPT-6 Sol, and only cached input fell, from $0.20 to $0.10. What actually dropped is the price of capability. Work that needed Astra on OpenAI's tests can now run on the cheaper model. A vendor whose summarising or document assistant was running on the expensive tier can swap down and, if OpenAI's numbers hold, keep most of the quality. Their cost per task falls. Your invoice stays where it was. The cached-input cut matters more than it looks, too, because enterprise add-ons tend to stuff the same CRM record, contract or policy manual into request after request. That is exactly the traffic that gets cheaper.

Now look at what the vendors charge. Microsoft 365 Copilot launched as a $30 per user per month add-on on top of an existing Microsoft 365 licence. Microsoft has since cut the business rate and layered on promotions, and the pricing guides cannot agree on today's figure. When nobody can say what a product costs, I read that as a price that was never tied to cost in the first place. Salesforce runs three meters for Agentforce at the same time: $2 per conversation, Flex Credits at $0.10 per action sold in packs of 100,000 credits for $500, and digital labour licences from $125 per user per month. Three meters for one product tells me Salesforce is still testing what buyers will pay, not pricing from its own costs.

Google has already shown where this goes. It stopped selling Gemini as a $20 to $30 per user add-on and folded it into Workspace base plans, comparing the old Business Standard plus Gemini at $32 with the new Business Standard at $14. One analysis calls the move bundling and segmentation rather than a clean cut, and what you pay depends on how your account was set up. Fair enough. Google still decided that charging separately for AI was a line item it could no longer defend, and it made that call before this latest drop in model prices.

Vendors have a decent reply. They say they sell integration, permissions, data plumbing and support, and none of that got cheaper. They can point out that OpenAI itself says Astra, at 68.1%, remains the model for the hardest research. They can note that OpenAI is launching Ultrafast, a premium tier running up to 8x faster in Codex and 6x in the API, so speed still commands a price. And they can argue frontier costs may not fall further, since the Wall Street Journal reported that OpenAI scrapped the expected GPT-6.1 Astra after internal testers raised safety concerns, including more deception.

I accept all of that and still side with the buyer. Integration had value before the AI arrived, and you already pay for it in the base licence. The add-on is the AI part, and the AI part is what got cheaper. Most of what a sales rep or finance clerk asks of these assistants is questions about documents and multi-step routine workflows. Those are the tasks where OpenAI claims Sol beats Opus 5.5 on its GDP.pdf test at under half the cost per task, and on AutomationBench at roughly a third. If your vendor insists that summarising call notes needs an Astra-class model, get that claim in writing. If you need speed, pay for speed where you need it, as its own priced option, and stop letting it justify a flat per-seat premium across the whole company.

Before your next renewal, ask each vendor which model runs the feature you pay extra for and what their cost per task was a year ago compared with now. Then divide Agentforce's $0.10 per action by Sol's $2 per million input tokens, and put the number of tokens that dime buys in the first line of your negotiation email.

Prompted by Introducing GPT-6.1 Sol, OpenAI.