Custom Software vs. SaaS in 2026, Now That AI Has Moved Both Sides of the Math

Something changed in SaaS pricing in 2026 that has nothing to do with the usual annual price hike. Growth Unhinged's survey of 230 B2B software and AI companies, run in April and May of this year, found that hybrid pricing, mixing a subscription with usage or credits on top, jumped from 25 percent of companies to 37 percent in twelve months. Twenty-nine percent already sell AI credits outright, and roughly a third of the rest plan to add them within six to twelve months. The reason is margin, not greed: median gross margin on AI-powered features sits near 50 percent, well under the 70 to 80 percent that funded generous flat-rate plans for a decade. Seventy percent of buyers in the same survey said the AI spend comes out of their existing software budget rather than new money, whether they asked for the feature or not.

That is a structural shift, not a sticker-price bump. A flat per-seat bill is something you can underwrite for three to five years. A usage meter is not, by definition. It moves with how hard your team leans on the AI feature, which is exactly the part of your bill you do not control.

Here is where I have to steelman my own side of this argument, because the other half of the equation, AI making custom builds cheaper, is weaker than the marketing suggests. DX tracked engineering velocity across more than 400 companies over 14 months and found a median pull request throughput gain of just 7.76 percent, with most companies landing between 5 and 15 percent, even as AI tool usage rose 65 percent in the same period. Only the top 10 percent of companies saw anything close to the "AI writes most of the code" story, at 43.9 percent. And that modest gain is not free: DX puts a full AI coding stack, an inline tool like Cursor plus an agentic one like Claude Code, at $200 to $600 a month per engineer, a real new cost line. Coding itself is only about 14 percent of a developer's workday, which caps how much any of this can compress total build cost. If you are penciling in a custom build assuming AI cuts your engineering bill in half, the actual data says be more conservative than that.

A modest gain is still a real one, though, and it compounds in the opposite direction from where SaaS pricing is heading. A 7 to 15 percent throughput gain lowers the cost of the side of the comparison you control. A shift toward usage-based AI billing raises the volatility, if not always the average cost, of the side you do not. Add those together and the case for owning a narrow, custom tool no longer rests only on "the vendor keeps raising the sticker price," the argument that has always applied. It now also rests on "the pricing model itself is becoming a moving target," which is a much harder thing to plan a five-year budget around, even when the average bill looks similar to last year's.

Before you renew a narrow, vertical SaaS tool this year, ask the vendor one direct question: is the AI feature a fixed add-on, or a usage meter that scales with how much your team relies on it? A fixed add-on is annoying but budgetable. A meter means your cost can move independent of your headcount, your usage patterns, and your control, which is precisely the kind of unpredictable bill worth re-pricing against a scoped, owned alternative, especially now that your own engineers are probably already getting some real, if modest, throughput gain from their own AI tools.

This is the comparison our custom software team runs with current numbers, not last year's, because both sides of it are moving faster than they used to. If a renewal just showed up with a new usage line you cannot forecast, talk to us before you sign another multi-year term on a price nobody can actually predict anymore.

Related reading

• OpenAI Is Now Selling the Agent Plumbing Most Engineering Teams Were About to Build
• What a Forward-Deployed Engineer Does, and When You Need One Instead of Another SaaS Seat