AI models got 80% cheaper this summer. The tools built on them didn't.

The cost of running an AI model collapsed this summer. The cost of the AI tools you actually buy did not move. On July 30, 2026 OpenAI cut GPT-5.6 Luna by 80%, and DeepSeek came out of preview cheaper still — but at the product layer, 65.1% of the active AI tools we track still require payment and only 5.0% are genuinely free. AI tool pricing has decoupled from the price of the intelligence underneath it. This is what our catalogue shows, as of August 13, 2026, and where the savings did land.

The short answer

Model prices are in a price war. Application prices are not, and there is no particular reason to expect them to follow.

Of the 2,457 active tools in our directory, 1,438 (58.5%) are paid outright, 735 (29.9%) are freemium, 162 (6.6%) are free trials and 122 (5.0%) are free. Two-thirds require money to use in earnest. That distribution is what it is after a year in which the marginal cost of a token fell by roughly an order of magnitude.

Where cheaper inference has shown up is not evenly spread. It is concentrated almost entirely in categories where the buyer can walk away.

What actually happened to model prices

On July 30, 2026, three weeks after launching the GPT-5.6 family, OpenAI cut its prices:

• GPT-5.6 Luna: $1 / $6 per million input/output tokens → $0.20 / $1.20, an 80% cut
• GPT-5.6 Terra: $2.50 / $15 → $2 / $12, a 20% cut
• GPT-5.6 Sol, the flagship: unchanged at $5 / $30

The pressure is visible in what was cut and what was not. The cheap, high-volume tier got the 80% reduction; the frontier tier did not move. That is the shape of a company defending the commodity end of its range while protecting the premium end.

The competition explains why. DeepSeek V4 Flash left preview this month at $0.14 / $0.28, and V4 Pro sits at $0.435 / $0.87 with a standing promotional discount. CNBC reported in early July that Chinese models had taken 46% of US enterprise token usage on OpenRouter, at times running ahead of US-origin models. Luna at $0.20 undercuts DeepSeek on input specifically.

For anyone building on these APIs, that is a genuine and large cost reduction. The question this post is about is whether any of it reaches the person buying the finished tool.

What that did to what you pay: nothing visible

Here is the whole active catalogue by pricing model, as of August 13, 2026:

Two ways to read it, both true. 65.1% of active AI tools require payment — paid outright or time-limited trial. And just one in twenty is free in the sense of costing nothing to keep using.

We are not claiming tool prices rose. We are making the narrower point that the input-cost collapse is not visible in how AI products are sold. The pricing model is the structural commitment a vendor makes to its buyers, and across two and a half thousand products it still overwhelmingly says "pay us".

Where the savings did land

The interesting part is the spread. Across the 23 categories with at least 50 active tools, the share that is free or freemium runs from 84.2% down to 14.4% — a six-fold difference.

The pattern is not about how expensive the software is to run. Running a security platform is not six times more compute-intensive than running an MCP server. It is about who is on the other side of the table.

At the top of that table sit categories where the buyer is an individual or a developer who can leave on a Tuesday afternoon: MCP servers, note-taking, image generation, learning tools. Free tiers there are customer acquisition, and cheaper inference makes that acquisition cheaper to fund — which is exactly where you would expect a cost saving to surface first. We wrote about that category's growth in the state of MCP servers.

At the bottom sit categories bought by committees, with procurement, security review, and integration work measured in months: security, healthcare, analytics, sales, support. Nobody in those categories competes on a free tier, because a free tier does not shorten a six-month enterprise sales cycle. Cheaper tokens improve the vendor's margin and change nothing about the price on the contract.

Why the discount stops at the application layer

Four structural reasons, none of which require anyone to be behaving badly.

Inference is a minority of the cost of running a software company. Engineering, sales, support, compliance and infrastructure do not get cheaper because a token does. An 80% cut on a line item that is a fraction of cost of goods sold moves the total very little.

Software is priced on value, not on cost. A tool that saves an analyst six hours a week is priced against those six hours. That number is unchanged by OpenAI's price list, and no vendor cuts its price because its supplier did unless a competitor forces it.

Most AI tools do not sell tokens, they sell seats. The unit a buyer pays for is a person or a workspace. Input costs falling shows up as improved gross margin inside that seat price — which is, incidentally, one of the few genuinely good pieces of news for the survival question we looked at in when will the AI bubble burst.

Switching costs absorb the difference. The steeper the switching cost, the less the price has to move. That is the mechanism behind the six-fold spread above, and it is the same mechanism that killed a standalone AI browser while agentic browsing itself carried on.

What we deliberately did not measure

We can tell you what AI tool pricing looks like today. We cannot honestly tell you how it has changed, and it is worth saying why, because the chart would have been the most shareable thing in this post.

Our records carry the date a tool entered our catalogue, not the date it launched or last changed its pricing. Slicing pricing by that date produces a dramatic-looking trend — and it is an artifact of when our editors ran sourcing batches, not of anything happening in the market. Same for survival: paid tools appear to fail more often than freemium ones, but the paid-heavy entries are also the oldest, so age and pricing model are tangled together and the comparison cannot carry weight.

So this is a snapshot, stated as a snapshot. Both halves of every ratio come from the same catalogue on the same day. Anyone wanting the trend will have to wait for us to measure it prospectively, which we can now do because this post fixes today's number in place.

What this means if you are buying AI tools in 2026

Do not wait for prices to fall. Nothing in this data suggests application prices track model prices. If a tool is worth its price today, the API price list is not a reason to delay.

Use the model price collapse as a negotiating fact, not a hope. If a vendor's product is a thin layer over an API whose price fell 80% in a quarter, that is a fair thing to raise at renewal. It works best where switching costs are low — which is precisely where free tiers already exist.

Treat a free tier as a signal about the market, not about generosity. Free is concentrated where buyers have power. A category with no free tiers anywhere is telling you switching is hard, and you should price the exit before you sign.

Check what the price includes as models get cheaper. Cheaper inference makes it economical for vendors to widen limits rather than cut prices. More generous usage at a flat price is a real saving, and it will not show up as a smaller number on the invoice.

How we know

Every figure here comes from our own catalogue, where each tool is reviewed by an editor and assigned one of four pricing models — free, freemium, free trial, or paid — from the vendor's own published pricing. Both halves of every ratio come from that same catalogue on the same day, so each rate holds regardless of how complete our coverage of the market is, but it describes the tools ToolDirectory.AI tracks rather than the whole AI market.

Only active tools are counted, because a shut-down product has no current price; the retired ones live in our AI graveyard. Category rates use a 50-tool floor, because a percentage over a dozen products is noise rather than a signal. Counts are given with their denominators throughout, and figures are as of August 13, 2026.

Model prices are from OpenAI's own July 30, 2026 announcement and contemporaneous reporting.

Frequently asked questions

Are AI tools getting cheaper in 2026? Not at the product layer. As of August 13, 2026, 65.1% of the 2,457 active AI tools we track require payment and only 5.0% are free, even though model API prices fell sharply this summer. The savings so far are visible in model pricing, not in what finished tools charge.

How much did AI model prices fall in 2026? OpenAI cut GPT-5.6 Luna by 80% on July 30, 2026, from $1/$6 to $0.20/$1.20 per million input/output tokens, and Terra by 20%. The flagship Sol tier was left unchanged at $5/$30. DeepSeek V4 Flash left preview this month at $0.14/$0.28.

Why are AI tools still expensive if AI got cheaper? Inference is a minority of what it costs to run a software business, software is priced on the value it delivers rather than on cost, most tools sell seats rather than tokens, and switching costs mean vendors are rarely forced to pass savings on.

What percentage of AI tools are free? 5.0% of the active tools in our catalogue are free, and a further 29.9% are freemium, so 34.9% can be used at no cost at some level. The remaining 65.1% are paid or time-limited trials.

Which AI tools are most likely to be free? MCP servers, at 84.2% free or freemium, followed by productivity tools at 66.0% and education tools at 62.8%. Free tiers cluster where buyers are individuals or developers who can switch easily.

Which AI categories almost never have a free tier? Security and governance at 14.4% free or freemium, healthcare at 17.0%, BI and analytics at 17.9%, and sales tools at 18.9%. These are committee purchases with long sales cycles, where a free tier does not help the vendor.

Will AI tool prices come down as models keep getting cheaper? There is no evidence of it so far. The more likely form of the benefit is wider usage limits at the same price rather than lower prices, and better vendor margins — which matters more for whether a tool survives than for what it costs.

Where to go next

The spending side of this story is in the AI capex bubble, and the survival side in when will the AI bubble burst and is the AI bubble bursting?. For risk by category rather than price by category, see AI tool failure rate by category. And before you pay for anything, how to tell if an AI tool is legit.

— The ToolDirectory.AI editorial team