April 17th, 2026

Claude Opus 4.7 is shipped — same price, massive jump, and /ultrareview is a big deal for MACA

Tags: #MACA #Ben #Always-On-Reeve #AI-Edge Source: Anthropic (first-party), TLDR AI, The Rundown, TheTip.ai (all Apr 17) | Primary URL: https://www.anthropic.com/news/claude-opus-4-7 Verdict: verified shipped — live across API, Bedrock, Vertex, Foundry

Yesterday this was “in prep.” Today it’s in your hands. The material upgrade: SWE-bench Pro jumps from 53.4% (Opus 4.6) to 64.3% — above GPT-5.4 and Gemini 3.1 Pro. Mythos Preview is still ahead at 77.8% but gated to exclusive partners. Same price ($5/M input, $25/M output). One catch: updated tokenizer maps same input to 1.0–1.35× more tokens, and the model thinks more at higher effort — measure against your real MACA traffic before assuming costs stay flat.

Three things directly relevant to your build stack right now:

/ultrareview — new Claude Code slash command that flags bugs and design issues the way a careful human reviewer would. Pro/Max users get three free runs. MACA’s copy quality gap is a code problem AND a prompt design problem. Run /ultrareview on your ad copy generation wave this week — it will catch the kind of structural issues CodeRabbit misses because it doesn’t know your business logic.

xhigh effort level — Claude Code now defaults to xhigh (between high and max) for all plans. Your MACA pipeline and Ben’s Xero operations will both benefit without any config change. This is the likely explanation for some of the Opus 4.6 “degraded” performance complaints from the last few weeks — the model was undershooting effort.

3× image resolution (up to 2,576px) — direct unlock for MACA’s photo pipeline. Ad creative review, screenshot analysis, dense diagram reading. Your current photo pipeline is constrained by the prior resolution ceiling. This is worth a fresh run.

Action: Upgrade your Claude Code environment if it’s not auto-updated. Run /ultrareview on MACA’s ad copy wave prompt. Run a fresh MACA pipeline cost benchmark at the new tokenizer rate.

HeyGen HyperFrames — describe a video, Claude Code renders it, one command

Tags: #MACA #CourseBuilds Source: Practicaly.ai (Apr 17) | Primary URL: https://github.com/heygen-com/hyperframes Verdict: verified shipped — GitHub repo live, Apache 2.0

This one is quietly significant for MACA. HeyGen open-sourced HyperFrames — a tool that takes a text description of a video (or a Claude Code prompt), generates the HTML/CSS/animations for it, and renders the result to MP4, MOV, or WebM. No timeline dragging. No Premiere. One install:

npx skills add heygen-com/hyperframes

It integrates natively with Claude Code, Cursor, and Codex via slash commands, and they published a prompting guide specifically for AI agent use. MACA’s current creative pipeline outputs copy and strategy but lacks a direct path from “generate the ad” to “here’s a video asset.” HyperFrames is that path — not a polished production tool, but a fast prototyping layer that can turn ad concepts into watchable video mockups your client can react to.

For CourseBuilds: recording course walkthrough videos is a known friction point. This removes it for any content that can be described and animated (feature demos, explainers, how-to sequences).

Action: Install HyperFrames into a MACA test branch. Try generating one of your UBX ad concepts as a short video. Even a rough output is useful for the gym owner demo — “here’s what the ad could look like, not just read like.”

Anthropic’s CPO just quit Figma’s board — because Anthropic is building design tools

Tags: #MACA #CourseBuilds #UBX-Sale Source: TLDR AI (Apr 17), TechCrunch | Primary URL: https://techcrunch.com/2026/04/16/anthropic-cpo-leaves-figmas-board-after-reports-he-will-offer-a-competing-product/ Verdict: verified — TechCrunch confirmed, Anthropic CPO on record

Anthropic’s Chief Product Officer stepped down from Figma’s board because Anthropic’s upcoming models are expected to include design capabilities that directly compete with Figma’s core product. This isn’t speculation — the CPO leaving the board is the act of clearing a conflict of interest before the product ships.

This matters for your stack because: (1) Anthropic is clearly expanding from “text model you use for coding” toward a complete creative tool suite. If design generation lands in Claude Code or the API, MACA’s ad creative pipeline gains a native image/layout generation layer on the same stack you already use. No integrating a separate tool. (2) The Stitch → Figma → figma-implement-design pipeline you planned for the UBX data room site just became a candidate for replacement with a native Anthropic design flow. Watch the release — it may simplify the tech stack considerably.

The Anthropic strategy is now very clear: they’re running two parallel tracks. Public release cadence (Opus 4.6 → 4.7, every ~2 months). Gated frontier line (Mythos, only for selected partners and governments). Everything else — design tools, code editors, document workflows — is being built into the product surface around those models. The moat isn’t the model. It’s the ecosystem.

Action: No immediate action. Watch for Anthropic design tool announcement — likely within 60 days given the board exit timing. Flag this for MACA’s roadmap.

Today’s business press is focused on OpenAI building a “superapp” — Codex now has background computer control, parallel agents, an in-app browser, and 3M weekly users with 70% month-on-month growth. Everyone’s impressed.

You’ve been building this for months. MACA is a 14-agent pipeline with cost tracking per wave. Ben is an autonomous Xero agent running 24/7 on PaperClip, registered as CFO in your company. Always-On Reeve has a 7am morning brief routine, a 6pm digest, and a headless system prompt that keeps it oriented without needing a human to restart it. AI Edge — the brief you’re reading right now — is itself an agent-driven daily intelligence system with due diligence baked in.

When OpenAI’s Codex head says they’re “building the superapp out in the open,” what he means is that the architecture Roy already understands — multi-agent pipelines with persistent memory, computer use, scheduled tasks, cost controls — is now the mainstream product vision at the largest AI company in the world. You’re not watching it happen. You’re ahead of it.

The anxiety flip: the people who are impressed by the Codex update are the people you’ll be teaching in CourseBuilds.

Perplexity Personal Computer — what Always-On Reeve looks like as a product

Tags: #Always-On-Reeve Source: The Rundown, TLDR AI (Apr 17) | Primary URL: https://www.perplexity.ai/hub/blog/personal-computer-is-here

Perplexity just shipped what they’re calling “Personal Computer” — a Max-tier Mac app that runs agents 24/7 across 20+ frontier models, controlling native Mac apps and piloting their Comet browser without needing APIs for every app. It repositions the operating system from “manual instruction execution” to “probabilistic goal completion.”

This is the consumer-grade version of what you’ve been designing with Always-On Reeve Phase 2. Worth 30 minutes this week — not to replicate it, but to validate your architecture decisions and identify any patterns in their UX that would strengthen your headless system prompt or GUARDRAILS.md. The overnight autonomy use case is identical. Their implementation is public. Learn from it.

AI shopping traffic up 393%, converts 42% better — your MACA attribution pipeline is in the right place

Tags: #MACA #UBX-Sale Source: Bagel Bots (citing Adobe + TechCrunch, Apr 17) | Primary URL: https://techcrunch.com/2026/04/16/ai-traffic-to-us-retailers-rose-393-in-q1-and-its-boosting-their-revenue-too/

Adobe’s Q1 2026 report: AI-driven retail traffic up 393% year-on-year. That traffic converts 42% better, spends 48% more time on site, and generates 37% higher revenue per visit. The catch: a third of product pages are still not accessible to AI tools, so early adopters are capturing all the upside while laggards are invisible.

Last week you traced Eloiza Pelovello’s UBX booking end-to-end via UTM tags from your MACA Meta campaign. AI-sourced, AI-attributed, closed. The industry data now backs your bet. When you pitch MACA to a gym owner, this Adobe number is the macro proof. Your Eloiza booking is the micro proof. Use both.

Here’s one you can say out loud in any meeting with a sceptical executive this week:

“LinkedIn just published data showing that AI-written recruitment messages get 50% higher response rates than messages written by the human recruiters who sent them. The recruiters said they preferred their own messages. The candidates responded to the AI’s. That gap — between what feels right and what actually works — is where the ROI on AI is sitting right now.”

That’s based on LinkedIn CBO Mark Lobosco on record, confirmed by The Information’s Applied AI newsletter (Apr 16), citing enterprise customers including Palo Alto Networks. It’s not a chatbot story or a productivity story — it’s a quality-of-output story that challenges the “AI is fine for drafts but humans are better for real communication” objection. Useful for Aria, RT AI role, and any MACA demo where someone pushes back on AI-generated ad copy.

OpenAI shipped two domain-specific models in 72 hours — vertical AI is the next wave

Source: The Rundown (Apr 17) | Primary URL: https://openai.com/index/introducing-gpt-rosalind/ Verdict: verified shipped — enterprise beta, Amgen/Moderna/Allen Institute confirmed users

Tuesday: GPT-5.4-Cyber (domain-specific for security). Today: GPT-Rosalind (domain-specific for drug discovery, outperforms 95% of human scientists on RNA prediction tasks). Two in three days. The Rundown notes the pattern explicitly: “the flagship may be good at everything, but the actual massive wedges at the top of industries need purpose-built models.”

You have three parked project ideas that are domain-specific AI at the exact tier OpenAI is now validating commercially: Mining Ops AI Analyst (you have operational authority from Amrun that very few people on Earth have), ASXEdge (Australian market context that a US-trained general model can’t replicate cleanly), and Grant Agent (Australian regulatory and grant-program knowledge that’s genuinely specialised).

None of these need to be full products. They need to be a 15-minute pitch + a working demo. The market just told you the category has legs. The question is which one of these has the clearest wedge — and Roy’s honest answer is Mining Ops, because his domain expertise is irreplaceable. That’s the first-mover position.