Read the two sentences most people skimmed past in Figma's launch note for the Make properties panel. They are not about spacing or typography. They are about credits.
"Direct manipulation gives Make a precise target, instead of the searching and guesswork a text prompt requires. Each edit uses far fewer tokens and finishes faster than describing the same change in a prompt."
And then the detail that gives the game away: "Edits stack up in the chat panel first, staged and credit-free, until you apply them."
The properties panel is a token-reduction engine wearing the costume of designer familiarity. Once you see it that way, the whole convergence war reorganizes around generation cost.
Everyone read the properties panel as a comfort feature—a return of familiar Figma controls for anxious designers who don't want to type into a chat box. That reading is wrong, or at least it buries the lede.
The properties panel is a token-reduction engine wearing the costume of designer familiarity.
Once you see it that way, the whole convergence war reorganizes around a single scarce resource, and it isn't talent, or taste, or feature count. It's generation cost.
Here is the thing the "design tools add AI, AI tools add design" framing gets backwards. People describe this as two armies marching toward the same midpoint and colliding. It's not a collision. It's a race to control the cost of a single generated token.
Consider what a text prompt actually is, economically. When you type "make the padding a little tighter and bump the heading weight," the model has to locate the element, infer your intent, generate a candidate, and often iterate when it guesses wrong. Every one of those steps burns compute, and every miss burns it again.
Now consider dragging a padding value from 24 to 16. There is no searching. There is no guessing. The target is exact, and the instruction is unambiguous.
That is the sentence Figma buried: fewer tokens, faster finish. A properties panel is a lossless compression format for intent. It hands the model coordinates instead of a description, and coordinates are cheap.
Watch the staging mechanic, because that's where the strategy hardens into product. Edits accumulate credit-free in the chat panel until you commit them. You can try three corner radii and four font weights, dragging elements around until they "feel right," and none of it costs a thing until you apply.
Figma has effectively decoupled exploration from billing. The expensive part—generation—only fires on commit, and by then the instruction has been compressed to something the model can execute in one pass.
That is not a familiarity feature. That is a pricing architecture.
Now widen the frame, because everyone else is playing the same game from different starting points, and they're all optimizing the same variable.
Framer said the quiet part with a discount. When OpenAI cut prices, Framer announced that Terra and Luna "now use even fewer credits," and stacked that on top of a 67% efficiency improvement they'd landed that month, plus both models at 50% off through August 14, per Framer.
Read that stack again. Four separate levers—vendor price cut, in-house efficiency gain, model-level margin, and a promotional discount—all pull in one direction: drive down the cost of a generation until trying Framer feels free.
Framer isn't marketing features. It's marketing unit economics. And its users have noticed the same thing Figma noticed: one Framer thread openly wishes for "a lightweight AI for those everyday tasks"—spacing, typography, simple edits—instead of firing "the latest AI model just to fix spacing." That's a user asking for cheaper steering, in exactly the language the properties panel answers.
Then there's Claude Design, which arrived at the same destination from the opposite direction and didn't even mean to.
Nate Parrott, the lone designer on the Claude Code VS Code project, built it as a side project after Opus 4.5 shipped in November 2025 and his two engineers suddenly started outpacing him. His first instinct was the obvious one: pack up screenshots and a requirement and ask Claude to "design it for us." The results were bad.
The breakthrough was a pivot in what he was asking the model to produce. He stopped asking Claude to draw designs and started asking it to generate HTML.
That is not a cosmetic change. It is a change in the unit of generation.
HTML, Parrott realized, is a rich interactive medium—slides, prototypes, animations, and, per UX Planet, reportedly even 3D objects all flow through the same pipeline. By making HTML the universal canvas, he gave the model a target format that's precise, editable, and cheap to regenerate in pieces rather than redraw wholesale.
A million people used it in its first week. Not because it draws prettier pictures, but because it turned generation into something structured enough to steer.
And notice what Parrott did next, because it rhymes exactly with Figma's staging move. He distilled Anthropic's brand—fonts, colors, assets, design principles—into the prompt system, so every generation starts pre-constrained. Claude Design now scans your codebase and design files to build a system automatically, then applies your colors, fonts, and components to everything.
That's the same logic as Figma's properties panel using the tokens already defined in your codebase. Constrain the generation up front, and you spend fewer tokens correcting it after.
Gamma is the control group—the tool that shows what happens when you optimize generation without owning the steering layer.
Gamma is fast precisely because it makes the design decisions for you. Give it a topic, get a deck. But as one UX Planet teardown lays out plainly, that's also its ceiling: no pixel-perfect screens, no detailed annotations, no custom data viz, no reusable design-system components. Its own docs admit advanced image editing isn't supported in the editor.
Gamma cheapened generation by removing precise control. Figma, Framer, and Claude Design are cheapening generation by adding it. That is the whole argument in one contrast.
So the "design-first versus AI-first" framing collapses. Nobody is fighting over whether to lead with design or lead with AI. They're fighting over who owns the cheapest, most precise mechanism for steering a model—because that mechanism is what makes AI code generation economically survivable at scale.
Prompting loses on this axis. It's expressive but lossy, and lossy means retries, and retries means credits. Direct manipulation wins because a dragged value is a compressed, exact instruction that lands in one pass. Figma's own note says as much: annotations exist for the cases that "span beyond visual properties," where you genuinely need words. Everything else, you point at.
The properties panel is the sharper strategic move not because it's more familiar, but because it's the most efficient encoding of intent yet shipped into an AI pipeline. It sells steering, and steering is the scarce good.
Which brings us to the tell nobody clocked: Config 2026.
Figma's brand team built the conference identity around "digital tools, human expression"—wonky, idiosyncratic, hand-drawn textures that also feel "programmatic and clearly digital." Read as brand strategy, it looks like a hedge against commoditization, a way of insisting there's still something human left when the machines can generate anything.
But look at how they made it. They built their own tools in Figma Make to codify three specific lo-fi textures—scribbly linework, blurred gradients, oval particles—then fed images through them to get controlled, repeatable results. They built a dithering tool to treat hundreds of headshots consistently without hours of manual work.
That is not a defense against AI generation. It is a demonstration of the exact thesis.
The distinctive output—the "opinionated in a way that feels unique to Figma" look—didn't come from raw generation. It came from purpose-built controls that steered generation toward a precise, idiosyncratic target. Custom sliders for wonkiness, in effect.
Which is what a properties panel is: a set of controls that convert human intent into cheap, exact instructions a model can execute without guessing.
Figma isn't hedging against commoditization with human touch. It's showing, in its own brand work, that the value has moved from generation to steering—and then selling the steering layer as a product.
The convergence war won't be won by whoever generates the most. Generation is already commodity, already 50% off through August 14. It'll be won by whoever makes intent cheapest to express.
Figma's answer is a value you can drag. That's the sharpest move on the board, because it's the only one priced in tokens.