Marketing with AI Weekly Roundup: The Everyday Claude Gets Agentic, Ad Creative Drops to Pennies, and Fable 5 Comes Back from the Ban

Edition #41 - Sunday 5th July 2026

Jul 05, 2026

Welcome to issue #41 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

It's 2026: So why is your marketing still running one prompt at a time?

2025 marketing: prompting ChatGPT or Claude Chat - starting from scratch, every single time.
2026 marketing: delegating whole projects to Cowork - it works from your own files and does the job, while you approve.

Ready to leave the 2025 way behind? The Claude for Marketing Masterclass shows you how.

Master Claude for the everyday marketing work that’s actually on your desk.

Learn More About Claude for Marketers

A quieter week than Cannes - but a more useful one. Three things happened that change what the tools in your existing stack can do, without you paying a penny more.

Anthropic made its everyday model genuinely agentic and gave it to every Free and Pro user as the new default.

Google released image and video generation cheap enough to treat creative like a rounding error - and quietly wired it into NotebookLM, where it now turns your documents into vertical social videos for free.

And Fable 5, the frontier model the US government pulled from the market three weeks ago, is back for everyone - with a lesson attached about never hard-wiring your workflows to a single model.

Underneath the launches sits this week’s data point: new research says three-quarters of marketers now lose three-plus hours a week fixing AI output. Generation is solved. Trust is the bottleneck.

Below: the three stories to act on, two steal-this workflows (with the exact prompts), the tools worth your time, and the study that should reframe how you measure AI’s payoff.

Let’s get into it.

🔥 Top Stories

1. Claude Sonnet 5: the everyday model just learned to run your workflows — and it’s the new default

Anthropic released Claude Sonnet 5 (June 30) and made it the default model for every Free and Pro user the same day (it’s also available on Max, Team and Enterprise, in Claude Code, and via the API, Amazon Bedrock, Google Cloud and Microsoft Foundry). The pitch: “the most agentic Sonnet yet” — reliable multi-step planning, tool use (browsers, terminals, connectors) and self-correction that, a few months ago, needed an Opus-class model. In plain terms: it can run a chain of work — research → brief → draft → repurpose — rather than answering one prompt at a time, and recover when a step goes wrong. For API users there’s introductory pricing of $2 per million input / $10 per million output tokens through August 31 (then $3/$15).

Why it matters / what to do: The Claude you (and your whole team, including everyone on the free plan) already use just got materially better at exactly the work that used to fail — long multi-step workflows that needed babysitting. Two moves this week: re-run any Claude workflow you’d quietly given up on because it lost the thread halfway through — the ceiling has moved. And if anything in your stack calls Claude via the API, the cost of running agents at scale just dropped to roughly a fifth of Opus prices — worth a look at which automations suddenly make economic sense.

2. Google just made ad creative cost pennies: images in 4 seconds, video at $0.10 a second

Google shipped two new generative media models (June 30) in the Gemini app, Google AI Studio and the Gemini API. Nano Banana 2 Lite generates images in roughly 4 seconds at $0.034 per image — Google’s fastest, cheapest image model yet. Gemini Omni Flash turns a still image into video at $0.10 per second of output, editable in plain language (”slow the zoom, brighten the product”). If you need sharp, legible text inside the image — headlines, UI mockups, stat callouts — the higher-end Nano Banana Pro handles that and blends up to 14 reference images for brand consistency, at $0.134 per image.

Why it matters / what to do: This is the price of creative collapsing in real time. Twenty to thirty on-brand ad or social variations now cost less than a coffee, and animating a static hero image into a short clip costs about 50p — no video team, no production budget conversation. Try this week: take one existing static LinkedIn ad or graphic, generate five variations in the Gemini app, and turn the best one into a five-second motion asset with Omni Flash. The discipline shifts from producing creative to choosing and reviewing it — which is exactly where this week’s research (below) says the real cost now lives.

3. Fable 5 is back — globally, and with a lesson worth keeping

Three weeks after the US government ordered Anthropic’s newest frontier models off the market (our June 13 lead story), the ban is over. The US Commerce Department lifted the export controls on June 30, and Anthropic restored Claude Fable 5 to all users worldwide on July 1 — on Claude.ai, Claude Code and Cowork, with AWS, Google Cloud and Microsoft Foundry re-enabling “as quickly as possible.” The fix: a new safety classifier that blocks the reported jailbreak in over 99% of test cases (flagged requests get rerouted to Opus 4.8). The fine print for paying users: Fable 5 counts toward up to 50% of weekly limits on Pro/Max/Team through July 7, after which it moves to usage credits. The more restricted Mythos 5 was restored earlier (June 26) to roughly 100 approved US organisations under the “Project Glasswing” programme. And notably, Anthropic is now drafting a joint framework with Amazon, Microsoft and Google for assessing the severity of AI jailbreaks — the first real cross-vendor safety standard of its kind.

Why it matters / what to do: When we covered the ban, the takeaway was: keep a documented model fallback and never hard-wire one frontier model into your workflows. Three weeks later, that advice paid for itself — teams with a fallback shrugged; teams without one scrambled. Now the reverse move: if you parked any Fable 5 workflows in June, switch them back on and check the quality difference against your fallback. And diarise July 7 — that’s when Fable 5 stops drawing on your normal plan limits and starts consuming usage credits, which changes the economics of using it as a daily driver versus reserving it for the work that genuinely needs it.

📋 The Playbook

1. How to get AI doing most of your marketing work (skills → roles → tools)

Grace Leung’s new walkthrough is the clearest framework we’ve seen for the jump from “prompting AI” to “directing an AI team” — and it maps directly onto the skills-and-agents features in Claude. (Transparency note: her video is an Ahrefs partnership, demoed inside their new AI marketing agent, Agent A — but as she says herself, the three levels work in “Claude or any platform,” and we’ve written them tool-agnostically here.)

Level 1 — Package your proven workflows as skills. A skill is a workflow you’ve already validated, written down once, that the AI runs on command. Before building anything, map what you actually repeat. Her prompt, adapted:

“Here’s my day-to-day marketing work: [describe it]. Organise these jobs into a skill library grouped by marketing discipline, and give me the mapping.”

Then build (or import) skills for the top recurring jobs — she demos an AI-search visibility gap analysis that outputs the priority questions a brand loses in ChatGPT/Perplexity, builds a shareable dashboard, proposes the top three topics to close the gap, and re-runs itself monthly on a schedule. One cost tip from the video: run routine skills on a mid-tier model (Sonnet-class) — she found the results held up at a fraction of the cost of Opus-class models.

Level 2 — Define agent roles, so skills know when to fire. Skills are actions; they don’t know when to run. A role turns them into a focused worker. Define each role with six elements: the role itself, the input it takes, the skills/tools it uses, the workflow it follows, the output you want, and the guardrails it stays inside. Save each role as a playbook, then instruct the AI’s memory to load it whenever you tag the role. Her example chain: a content-strategist role reviews the gap analysis and produces a 90-day topic cluster board → a blog-writer role picks the “now”-priority topics and drafts to your content rules (question headings, tables, internal links, FAQ) straight into WordPress via a connector. Always with human review before anything ships.

Level 3 — Turn your best workflow into a one-click tool. At level 2, your workflows still live in the chat. At level 3, you package one into a simple app anyone on the team can run — inputs in, your standard of output back, no prompting skill required. Her example: a community-research skill extended into a tool that digs forums for real customer questions, proposes content angles, waits for your approval, then generates the finished carousel visuals. That’s your logic and quality bar, baked in and shareable.

Try this week: run the mapping prompt, pick your single most-repeated marketing job, and move it up one level — freeform prompting → skill, or skill → defined role. That’s the whole game: one rung at a time.

Watch here

2. Motion graphics with zero design skills (Claude Design, four tricks)

Paul Lipsky’s system for animated marketing graphics — social clips, animated stat cards, product demos — uses no code, no timeline editor, and comes down to four tricks. Setup: go to claude.ai/design (included in any Claude subscription) and select the Animation template in the prompt box before every prompt below.

Trick 1 — Start from a template’s code (the easiest win). Free motion-graphic template sites let you copy the code behind any animation you like. Paste that code into Claude Design with: “Use this template, but…” and list your changes (”change ‘chapters’ to ‘steps’, swap the brand colour to #00d4ff”). Because Claude sees exactly what you want, you skip the back-and-forth entirely — one prompt, done. You can even paste two templates and ask it to merge them into one sequence.

Trick 2 — Recreate any animation from screenshots. See a motion graphic you love with no template? Screenshot its first frame and last frame (plus one mid-frame if something significant happens), attach both, and describe the animation step by step plus your content changes. One lesson from his own iteration: describe behaviour, not adjectives — “make the progress bar steadily increase the entire time so it reaches full when the graphic ends” worked where “make it smooth” failed twice.

Trick 3 — Sync graphics to your talking-head video. Get a timestamped transcript of your video or audio (his free method: give Gemini the file with the prompt “transcribe this file SRT style”), paste the transcript into Claude Design, and attach brand assets or screenshots for styling. It generates on-brand motion graphics matched to what you’re saying. Refine by referencing timestamps: “around the 17-second mark, blur out the three trick names.”

Trick 4 — Fake a product screen-recording. Screenshot the web page or UI you want to animate, use Claude Design’s “grab web element” bookmarklet to copy the code of the exact on-page element, then describe the sequence: “Create a video of [product] being used. It should feel like a screen recording of a real user. Animation 1: zoom into the search box. Animation 2: type [query]…” Result: a polished demo clip of your product (or a workflow) with no screen-capture session.

Two finishing moves: the markup tool lets you click any element in the animation and describe a change to just that element. And for export — the official route (download the project archive, have Claude Cowork convert it to MP4) works but is slow; his honest shortcut is playing the animation full-screen and recording it with QuickTime. Nobody can tell.

Watch here

🛠️ New Tools & Features

1. NotebookLM now turns your documents into 60-second vertical videos — free for everyone

Google added Short Video Overviews to NotebookLM (June 30): feed it your sources — a report, a case study, a blog post, even a spreadsheet of results — and it generates a ~60-second vertical video with narration, motion graphics and text burned onto the frame for sound-off viewing. Unlike NotebookLM’s cinematic videos (paid plans only), shorts are rolling out to all users including the free tier — because they’re powered by the cheap Nano Banana 2 Lite model from Top Story #2. It works in the mobile app from day one. The marketing use is obvious: a repurposing pipeline from long-form content to LinkedIn/Reels/Shorts video with zero editing. Two honest caveats from early testing: every video carries a recognisable dotted background you can’t prompt away, and it treats your sources as third-party material — so it explains about the content rather than speaking as your brand. Fine for explainers; not yet a brand-voice channel.

2. Vmake consolidates video localisation into one workflow

Vmake Labs launched an AI Video Translator (June 29) that bundles what used to be five tools into one pass: translation, dubbing, voice matching (it clones the original speaker’s voice into the new language), lip sync, and enhancement up to 4K — across 14 languages. For B2B teams the narrow-but-real use case is localising an existing webinar, product demo or campaign video for other markets without re-shooting or re-recording. Worth knowing the company’s broader platform skews consumer/UGC — evaluate the translator on its own merits.

3. Claude in Chrome is now generally available

Anthropic moved its Claude-in-Chrome browser agent from beta to general availability (July 2) for paid plans. It’s Claude living in a side panel of your actual browser — able to read the page you’re on, click, fill forms and carry out multi-step tasks across sites, with permissions you control per site. Marketing uses that make immediate sense: competitor page teardowns while you browse (”compare this pricing page against the three tabs I have open”), pulling data out of dashboards that lack exports, and form-heavy admin. Start read-only — let it summarise and compare before you let it click.

4. Alli AI makes your WordPress site readable to AI crawlers

A practical answer-engine-optimisation (AEO — the practice of getting your content cited in AI answers) release: Alli AI launched a WordPress plugin that serves pre-rendered, server-side HTML to AI crawlers like ChatGPT, Perplexity and Claude, so every word of your content is parseable by the models deciding whether to cite you — without changing anything human visitors see. If your site runs on WordPress with heavy JavaScript (most modern themes), there’s a decent chance AI crawlers currently see far less of your content than you think. This is a one-plugin fix worth testing against a crawler-simulation check.

📊 Research and Data

1. The “revision tax”: 76% of marketers lose 3+ hours a week fixing AI output

Optimizely’s new global study (released June 30; fieldwork by research firm Savanta, May–June 2026; 2,003 marketing leaders across the US, UK, Germany, Sweden, Netherlands, Australia and UAE) puts a number on something every team feels: 76% of marketers spend at least three hours every week editing, fact-checking or correcting AI-generated output. The biggest single source of added work is fact-checking and hallucination review (48%), followed by time lost moving information between disconnected systems (40%). The report’s name for it — the “revision tax” — is exactly right: the time AI saves on generation is being quietly spent again on review and clean-up.

Why it matters / what to do: This reframes how you should measure AI’s payoff. The bottleneck is no longer drafting — it’s trust and plumbing. Two moves: first, actually measure your team’s revision tax for one week (hours spent fixing or verifying AI output — you can’t manage what you haven’t counted). Second, attack the two named culprits directly: build a lightweight fact-check step into the workflow (sources required with every draft, a standing verification pass) rather than treating review as invisible overtime, and connect the systems your AI pulls from so people stop re-keying context between tools. Note the week’s neat symmetry: creative generation got 10x cheaper (Top Story #2) while this study shows verification is where the hours go. Budget accordingly.

💡 Quick Hits

1. OpenAI previewed GPT-5.6

(June 26) in three variants — Sol, Terra and Luna — with access initially gated through a new US government review process; broad ChatGPT and API availability is promised “in the coming weeks.” No marketing-specific capability announced yet.

2. Anthropic launched Claude Science

(June 30), a research workbench for scientific analysis — niche for marketers, but a signal of how fast Claude is verticalising into profession-specific products.

3. ChatGPT Business added workspace plugin management

Admins can now discover, govern and control which plugins their team instals. Housekeeping, but useful if your marketing team runs ChatGPT at scale.

4. Claude for Enterprise added admin analytics and spend alerts

(July 2) — usage and cost dashboards by group and user, showing what’s actually being created alongside what it costs, plus model-level entitlements. The “prove the ROI of our AI spend” dashboard marketing leaders keep being asked for.

👀 On Our Radar

AI-search visibility is getting its own index.

Profound, one of the AEO-monitoring vendors — launched the Profound Index this week at its Zero-Click New York event: a standing benchmark scoring how often brands appear in AI-generated answers across the major assistants.

Last week we noted the AI-visibility tool category multiplying; an index is the next stage — visibility inside ChatGPT, Claude and Perplexity becoming a tracked, benchmarked KPI you can be compared on, the way domain authority became shorthand for SEO strength. Nothing to buy yet — but expect “what’s our AI visibility score?” to start appearing in leadership decks, and be ready with an answer for your own category.

🔮 The Big Picture

The constraint moved from making the work to trusting it.

Put the week together: Sonnet 5 makes capable multi-step AI free-tier standard. Google makes creative generation cost pennies. NotebookLM turns documents into videos for nothing. And against all that abundance, Optimizely’s data lands: three-quarters of marketers are paying a three-hour-a-week revision tax, driven by fact-checking and disconnected systems. Generation is now effectively solved and effectively free — which means the marketers who pull ahead in H2 2026 are the ones who build the verification and plumbing around AI: review steps that live inside the workflow, sources attached by default, systems connected so context doesn’t get re-keyed. The scarce skill isn’t prompting. It’s designing work you can trust without re-doing it.

The unit of AI work is now the agent, not the chat. Sonnet 5 was built explicitly for agentic work. Grace Leung’s framework — the week’s best playbook — is a ladder out of the chat window: workflows become skills, skills get roles, roles become tools your whole team runs. Even Claude in Chrome going GA is the same idea in a different surface: AI that does the task where the task lives, not a text box you visit.

If you’re still re-typing your best workflow into a chat every week, that’s the gap to close — package it once, and it compounds. Depth beats breadth; one system shipped beats ten tools sampled.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Discussion about this post

Ready for more?