Model Prices Fell 80 Percent. Most AI Software Bills Did Not Move.

Here is a fact that should worry anyone who assumed cheaper AI models automatically mean a cheaper AI bill: model prices have collapsed and most AI software prices have not moved at all.

On July 30, OpenAI cut the price of GPT-5.6 Luna by 80 percent, from $1 and $6 per million input and output tokens down to $0.20 and $1.20, and cut Terra by 20 percent, from $2.50/$15 to $2/$12. Sol, its flagship tier, held steady at $5/$30. The move came less than a month after GPT-5.6 launched, which tells you how much pressure OpenAI is under. A CNBC investigation cited in that same coverage found Chinese models had captured 46 percent of US enterprise token usage on OpenRouter, with DeepSeek V4 Pro pricing at $0.435/$0.87 per million tokens on promotion and Kimi K3 undercutting at $3/$15. Underlying inference costs have fallen roughly 80 percent since 2023 industry-wide, per the same reporting: GPT-4-class inference that ran $30 per million tokens two years ago now runs under $3. That is a genuine price war, and competition, not any regulator, is who is winning it for buyers.

Now the part that should give you pause. An analysis of 2,457 active AI tools published this week found 58.5 percent require payment outright, another 29.9 percent are freemium with a paid tier required for real use, and only 5 percent are genuinely free. Put those first two together and 65.1 percent of the tools tracked require payment to use seriously, as of August 13. That share barely resembles a market where the raw ingredient just got four fifths cheaper. And it is not uniform: free and freemium adoption ranges from 84.2 percent among MCP server tools down to 14.4 percent among security and governance tools, a sixfold spread the analysis ties to how easily a buyer can switch, not to what the tool costs to run.

Before concluding the AI tool market is broken, give the other side its due. Software has never been priced off the cost of its inputs, and there is no reason AI tools should start now. A CRM's price was never tied to the cost of the database underneath it. Inference is also a minority line item for most serious AI products once you add engineering, support, security, and the sales motion that gets a business to actually adopt the thing. If inference was 20 percent of a vendor's cost base and it fell 80 percent, that frees up roughly 16 points of margin, not grounds for an 80 percent price cut. And most AI tools were never reselling tokens to begin with. They sell a seat, a workflow, an outcome. A price war between suppliers does not obligate every business built on top of those suppliers to pass the savings through.

That argument holds up, and it is worth taking seriously rather than waving off as vendors gouging customers. But it does not contradict the free-market case here. It confirms it. Competition is compressing prices exactly where switching is easy and commoditized, MCP servers and thin utility wrappers, and barely touching categories where switching is expensive, like security and governance tools with deep integrations and audit trails behind them. That is what a functioning market looks like when it hits a wall of real switching costs. It is not evidence that regulation should step in and force AI software prices down. That would only entrench the vendors big enough to eat a mandated markdown and absorb compliance overhead, while pricing the leaner competitors who might actually undercut them out of the game.

So the fix is not a complaint to a regulator. It is arithmetic you run yourself. If a tool you pay $50 a seat for is mostly a thin layer over a model call that now costs its vendor pennies, that is a candidate to build, not rent, because building it yourself means you capture the token price war directly instead of subsidizing someone else's flat seat price. If the tool has real proprietary workflow, data, or integration depth behind it, the seat price was never really about tokens and haggling over inference cost misses the point. Anthropic's own Claude Sonnet 5 promotional rate reverts from $2/$10 to $3/$15 per million tokens on September 1, a reminder that pricing moves both directions and today's number is not a permanent fact either way.

This is the exact calculation my software team runs with clients: which AI line items are commodity inference wearing a subscription, and which ones are worth paying for because the vendor built something you would spend more than the subscription to replicate. If you are not sure which bucket your AI tools fall into, that is a conversation worth having before your renewal date, not after. Tell me what you are running and I will tell you honestly which of it is worth keeping.