Framer has been the website builder that designers actually respect for a few years now, the one that produces sites that don’t look like they came out of a template gallery. With Framer 3.0 (June 16, 2026) it also became an AI tool worth taking seriously, and it did it in a way most competitors haven’t: the agents work inside the canvas, not in a chat window off to the side.

New in August 2026: You now control how hard the agent thinks

On August 12 Framer added reasoning control to the agent model picker. You choose how much thinking a model does before it starts building: light reasoning for quick edits and refinements, higher reasoning when you want the agent to plan a whole page properly. It also added Fast mode for models that support it, including Opus 5, for noticeably quicker generations — at the cost of more tokens per task.

This is a more useful lever than it sounds. The single most common way to waste credits in Framer is asking a heavyweight reasoning model to nudge some padding. Light reasoning for small edits and high reasoning for structural work is the setting most people should be using, and until now you couldn’t choose. (Framer: Agents Reasoning and Fast Mode)

Two smaller items from the same window: Video Shaders (Aug 11) let shader effects accept video as well as images, starting with Chromatic Aberration and real-time editing. And Plugins 4.1 (Aug 4) brought branching and publishing into the plugin API, so third-party plugins can now create branches, merge them, publish a site, and deploy to a custom domain. That last one mostly matters if you use marketplace plugins, but it’s a sign Framer is treating branching as core infrastructure rather than a feature.

New in July 2026: Agent credits go a lot further

On July 21 Framer shipped a round of efficiency work that changes the one number founders actually feel: credit burn. Every model now completes tasks in 15–30% fewer steps, uses 30–40% fewer tokens, and saves another 10–15% through better caching. Framer also worked directly with OpenAI to cut OpenAI-model credit use by about 25% on multi-step builds.

The bigger shift is the default. GPT-5.6 Terra is now the default model, and Framer says it’s more capable than GPT-5.5, twice as fast, and uses 37.5% fewer credits. Stack the efficiency work on top and a multi-step build with Terra runs on 60–67% fewer credits than GPT-5.5 did when agents first launched. Sonnet 5 and Opus 4.8 get the same treatment and land around 48% cheaper.

What this means in practice: the “you’ll blow through your credits in one layout session” complaint we’ve had since launch is materially smaller now. It’s not gone, since 1,000 credits on Basic still isn’t a lot if agents are your main way of working, but the same budget now buys close to double the work it did in June. Source: Framer: Improved Agents Efficiency.

New in June 2026: Agents, branching, and cheaper seats

Framer 3.0 shipped three things that matter.

Agents on the canvas. Framer’s AI agents operate on the same project you do: the same pages, components, CMS entries, styles, SEO settings, and publish workflow. Ask an agent to build a pricing page, tidy up inconsistent styles across a site, rewrite the copy on a landing page, or bulk-edit CMS content, and the output lands as native Framer work you can then select, nudge, and refine by hand. This is the key difference from a Lovable or a v0: there’s no handoff moment where the AI gives you a thing and you’re stuck with it. The agent’s work is just… your file.

Branching. You can tell an agent to work on a branch, review the diff, compare against the live version, and publish only when you’re happy. For a marketing site that’s actually getting traffic, this is the feature that makes agents safe to use at all. Most no-code builders still let AI edit the live thing.

Pricing got simpler and cheaper. Basic is $10/mo (annual), Pro is $30/mo, and editor seats were cut from $40 to $20. Agent usage runs on AI credits: 1,000/mo on Basic, 3,000 on Pro. External agents (connecting your own Claude Code or Codex) are free during the preview. Sources: Framer 3.0 announcement, Framer: AI credits and agent pricing, BusinessWire.

Who it’s actually for

If your product is the thing you build in Lovable or Bolt, Framer is where the site around it lives: the landing page, the pricing page, the blog, the changelog. It’s very good at that, and the CMS is genuinely usable by a non-technical person after an afternoon.

It’s also the right answer if you care what your site looks like. This is the honest differentiator. Wix AI and Durable will get you a site faster; they will not get you a site anyone compliments. Framer’s default output has taste, and the agents inherit it because they’re constrained to the same components and styles your project already uses.

Where it falls down

The learning curve is real, and the AI doesn’t erase it. Framer’s canvas borrows heavily from Figma: stacks, breakpoints, constraints, layers. If you’ve never used a design tool, expect a genuinely confusing first week. The agents help you skip making things, but you still need to understand the model to fix what they made. A non-technical founder can absolutely learn it. They just won’t learn it in an hour, which is what the marketing implies.

It’s not an app builder. No auth, no real backend, no application logic, no complex data relationships. If you find yourself trying to build a logged-in dashboard in Framer, you’ve picked the wrong tool. Go to Lovable or Bolt and keep Framer for the marketing site.

AI credits still meter out faster than you’d expect. The July efficiency work roughly doubled what a given credit budget buys, but 1,000 credits on Basic still won’t last long if you spend whole sessions iterating on a layout with agents. Budget for Pro if agents are going to be your main way of working.

Bottom line

Framer is the best-looking website builder available to non-technical people, and 3.0 turned it into one of the more thoughtfully-designed AI tools in the category. Branching in particular is a level of restraint most builders skipped straight past. Take it for your marketing site and your blog. Don’t try to make it your product. And give yourself a real week to learn the canvas before you judge whether it’s working.

Was this helpful?