July 28th, 2026

Tier 1 — Claude Opus 5 is live and it changes your model-cost math

Anthropic launched Opus 5 Friday: claude-opus-5, 1M context, $5 in / $25 out, effort dial, Automatic Fallbacks beta, mid-task model switching. SOTA on agentic terminal coding, knowledge work, agentic search, computer use — beat Fable 5 on ARC-AGI-3 (30.2%, 3x next best), IMO 2026 (42/42), Frontier-Bench, GDPval-AA. Default on Max, strongest on Pro, API live today. Verdict: verified shipped — confirmed across four independent sources (TLDR, Rundown, Practicaly, The Tip) with matching numbers and Anthropic first-party blog. Action: covered in PAY ATTENTION — swap Ben’s default, benchmark one MACA run, hold if numbers don’t move.

Tier 1 — Anthropic just told you to gut your CLAUDE.md files

Published July 24: The New Rules of Context Engineering for Claude 5 generation models. Anthropic stripped 80%+ of Claude Code’s system prompt for Claude 5 models (Opus 5, Fable 5, Sonnet 5) with no measurable loss on coding evals. Newer models have the judgment; thick rulebooks now get in the way. Progressive disclosure, simple tool descriptions, auto-saved memories, HTML-artifact references replace elaborate CLAUDE.md files. Verdict: verified shipped (Anthropic first-party blog). Action this week: ruthless prune pass across ~/Reeve/reeve-headless.md, Ben’s Agent SDK prompt, MACA per-wave prompts, and the CLAUDE.md files in Fillarup + InvoiceGen. Measure one Reeve session and one MACA run before/after; if outputs hold, the token savings compound codebase-wide.

Tier 1 — Claude Cowork’s “Record a Skill” is the Reeve pattern, native

Anthropic shipped Record a Skill in Claude Cowork: open Cowork → click + → Record → do the task while narrating your steps → save reusable skill → attach it to a daily or weekly recurring task. Same loop you built into PaperClip + claude_local: Reeve’s Morning Brief (7am AEST) and EOD Digest (6pm AEST) already fire as scheduled routines. Verdict: verified shipped (Cowork feature, walkthrough guide published in Rundown today). Action this week: record two skills MACA actually needs — (1) post-campaign Meta ad audit against Mike Futia’s 6-dimension rubric (sitting in your inbox since Apr 15), (2) UBX South Bank data-room checklist pass. Both can graduate to recurring skills before Aug 1 if the recording holds up.

The MACA cost-dashboard is pointed exactly where the puck is going

You landed api/lib/costs.ts, public/cost-dashboard.html, and scripts/photo-costs.json in PR #10 back in April — per-run cost logger, per-photo cost tracking, live dashboard. Anthropic today ships Opus 5 with an effort dial + Automatic Fallbacks; the whole industry is converging on “model choice + effort + cost per task” as the primary API surface. Every other MACA-scale operator is scrambling to instrument their pipelines now. You already have the observability layer built — you can measure the Opus 5 impact on unit economics before most people even understand the question.

Reeve’s cron routines are the “Record + recurring task” pattern, six months early

Reeve’s PaperClip stack — com.paperclip.server launchd daemon, KeepAlive, Morning Brief 7am AEST cron, EOD Digest 6pm AEST cron, headless prompt at ~/Reeve/reeve-headless.md — is the exact loop Cowork’s Record a Skill + attach-to-recurring-task shipped as a native feature this week. You built it out of PaperClip + claude_local because Anthropic hadn’t shipped the primitive yet. Now they have. Decision is now available: keep the custom PaperClip stack for the hooks and control it gives you, or graduate the simpler routines (AI Edge is a candidate) into Cowork skills to shrink the operational surface. Either way — the pattern’s yours to make deliberately, not to catch up on.

Anthropic — The New Rules of Context Engineering for Claude 5 (6 min read)

https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models

Direct source for the “gut your CLAUDE.md” call in Section 1. The 80%+ removal figure with no measurable loss is the number that lets you make the same cut across Reeve, Ben, MACA, Fillarup, and InvoiceGen without guessing. Read once, run the prune pass in a single focused session — before you write any new prompts on Opus 5.

a16z — Lighthouse or Landgrab? How to Pick Your AI Sales Strategy (~30 min read)

https://www.a16z.news/p/lighthouse-or-landgrab-how-to-pick

Two enterprise GTM playbooks with clean examples. Lighthouse (category creation, marquee logos, six-figure ACVs, 3-6 month cycles): Harvey with Allen & Overy → $11B valuation; Hebbia into KKR and BlackRock. Landgrab (displacement, standardised product, speed): Decagon 0→8-figure ARR in 18 months, valuation tripled to $4.5B in six months; Stuut deploys in <1 week vs 6-18 months for SAP/Oracle AR. Direct application: UBX South Bank sale is Landgrab territory (operator-owner network, MJ referral pipeline, Aug 1 pressure — speed wins), CourseBuilds is a Lighthouse play (Aria is your Casper, high buyer exposure, proof travels in commercial real estate). The a16z frame gives you the second-order traps to design against — pilot purgatory, hostage to the logo, grabbing land you can’t hold.

“Anthropic dropped Opus 5 on Friday — same price as 4.8 at $5 in $25 out, so half the price of Fable 5, and it hit 30.2% on ARC-AGI-3, which is triple the next-best model. Fourth Anthropic model in under two months. And on the 24th they published that they’d deleted 80% of Claude Code’s system prompt for Claude 5 models with no measurable loss on evals — the era of thick prompt files is basically over.”

Use case: Any AI-pro conversation this week — RT AI role (R53597) touchpoint, the Aria pre-brief when the CourseBuilds Phase 0 slot opens with Michael Zaicek, or a UBX Australia call where “we’re building on the current frontier” matters. Signals you’re tracking pace of shipment (four models in two months), you know pricing curves cold, and you’re already adjusting stack accordingly (prompt trimming). Practitioner voice — not outside commentary.

Grok is now a free add-on inside Google Workspace — Aria’s wedge angle just shifted under you.

xAI shipped a free Grok add-on to the Google Workspace Marketplace on Friday. One install covers Docs, Sheets, Slides. Grok cites specific spreadsheet cells when answering, builds formulas and charts, drafts and refines documents in place, has Drive + email connectors. Aria (53 staff, all in Workspace, zero public AI footprint) is the exact target profile: giving every commercial leasing + Aria Living staffer a working AI assistant embedded in the tools they already use daily, at $0.

Act this week: install it in your own Workspace, run one lease-abstraction test in a Sheet + Doc to see how it feels vs the Claude-project approach you’ve been designing for Zaicek’s team. Decide: does the Grok add-on strengthen the Phase 0 wow (“here’s the free thing anyone can install, and here’s what a purpose-built Aria stack does that this can’t”) or weaken it? Answer is worth a 30-min test before the next Zaicek window opens — you don’t want to walk in unaware that half of Aria could install this tomorrow.

• [TLDR — “Prentis, Hoffman/Pincus computer-use lab, $100M at $1B”]: direction-of-travel, no stack overlap today.
• [TLDR — “Baseten fastest GLM-5.2 API, 280 tok/s”]: off-stack, no GLM anywhere in Prevail.
• [TLDR — “NVIDIA SANA-Video 2.0, 720p on single GPU”]: no video pipeline in your projects.
• [TLDR — “Celeris-1 diffusion LM, 15x faster”]: unvetted vendor claim, watch only.
• [TLDR — “NVIDIA ModelExpress P2P RDMA weight transfer”]: infra-layer, not your problem.
• [TLDR — “Brute Intelligence” (Benn Stancil)]: evergreen essay, not action.
• [TLDR — “NVIDIA Open Weights and American AI Leadership” PDF]: policy/geopolitics macro.
• [TLDR — “Legora BAR benchmark for legal AI”]: watchlist for CourseBuilds legal-play, nothing to do now.
• [TLDR — “Nanbeige4.2-3B / Laguna S 2.1”]: research releases, off-stack.
• [TLDR — “OpenRouter Classifiers beta”]: interesting for MACA long-term, tagged watch.
• [TLDR — “Prompt caching in agents” (earendil, 19 min)]: bookmark for when Ben’s cached-context economics matter.
• [TLDR — “Zvi: More on the OpenAI Hugging Face hack” (38 min)]: security post-mortem, worth reading in a security-hardening session, not today.
• [TLDR — “Perth PropelAuth MCP OAuth 2.1 deep dive”]: sponsor placement, not action.
• [Rundown — Multiverse CompactifAI sponsor]: vendor Sonnet-5-tier-at-65%-cheaper claim, needs independent vetting before trust.
• [Rundown — Roundtable AI use cases]: personal staff stories, no signal.
• [Rundown — “Apple AI glasses at WWDC 2027”]: two-year horizon.
• [Rundown — “Gemma 4 hits 300M downloads”]: Google DeepMind stat, macro.
• [Rundown — “Midjourney acquires Co-Star”]: astrology app, off-stack.
• [Rundown — “Meta AI agentic (Muse/Spark)”]: competitor infra, watch not act.
• [Rundown — Open-source letter (Nvidia/Meta/MS/47 others)]: covered in Section 1 context; the notable fact — Anthropic is the only U.S. frontier lab off the list — is the signal, but nothing to act on.
• [Rundown — Fin webinar / Merge Fusion / Photon-1 / OpenWorker]: aggregator quicks, no immediate stack fit.
• [TheTip — “Grok 4.6 and 4.7 announced”]: Musk noise same weekend as Opus 5, no ship details worth acting on.
• [TheTip — AI Money Group Skool $97/mo promo]: creator upsell, skip.
• [Practicaly — Heard macOS voice layer for Claude Code/Codex]: silent-unless-error voice UX — genuinely interesting for Always-On Reeve Phase 2 UX layer, park for that build.
• [Practicaly — Paradigm free-forever education platform]: off-focus for adult AI consulting.
• [Practicaly — Gamified CRM idea from roofing contractor]: fun frame, off-shape for current Prevail projects.
• [Info — “Startup OpenRouter fields multibillion takeover interest” (Palazzolo et al)]: routing plays consolidating — direction-of-travel for MACA model choice, article body needs fetch to be actionable.
• [Info — “DeepSeek pauses $7.4B round at $74B val”]: China AI finance, macro.
• [Info — “SpaceX Starship 13th flight since IPO”]: off-topic.
• [Info — “CXMT 472% Shanghai debut, $487B valuation”]: China memory chip macro.
• [Info — “Shein Hong Kong IPO filing, de minimis damage”]: e-commerce macro.
• [Info — “Nvidia $1B into Naver AI data center”]: sovereign AI infra macro.
• [Info — “Waymo exploring end of Uber partnership”]: AV infra, off-topic.
• [Info — “Nvidia $500B partnership with SK Hynix”]: chip-supply macro.
• [Info — “OpenAI cuts inference costs in half” (Palazzolo)]: Monthly Collection headline teaser only, body needs fetch, park.
• [Info — “Oracle data centers face multibillion cost surprise” (Vaughan)]: headline teaser only, body needs fetch.
• [Info — “Anthropic unusual post-IPO stock plan” (Weinberg)]: cap-structure story for later.
• [Info — “Google ‘Frozen’ chip”]: TPU-adjacent macro, headline teaser only.
• [Info — “Nvidia takes cut of customer cloud revenues”]: Nvidia business-model shift, macro.
• [Info — “Khosla-backed largest AI model on iPhone”]: edge-inference direction-of-travel.
• [Info — “Cursor reinventing itself as SpaceX deal looms”]: IDE-agent competitive read, no immediate stack impact.
• [Info — Monday Readout / Monthly Finance Collection / App Download promo]: curation emails, dupes above.
• [Info — “Nvidia priced-for-everything-to-go-wrong” (Ramaswamy)]: valuation commentary, macro.
• [Info — “AlphaSense tops $700M ARR, IPO steps”]: financial-research SaaS milestone, macro.
• [Info — “SoftBank/Altimeter/D1 into Thrive Holdings $2B”]: PE deal, off-focus.
• [Info — “Stargate power developer sells stake”]: infra macro.
• [Info — “Mercor’s fast growth relies on biggest AI cos”]: talent-marketplace story, macro.
• [Info — “Crackdown on AI Lovers in China”]: consumer-AI regulation, off-focus.
• [Bagelbots — “Prompt That Thinks Like a Startup Operator”]: another prompt template — you already have SOUL.md + voice-profile.
• [Neil Patel — SEO Week closes Monday, 5 months free]: course promo, skip.

• Sources scanned: 13 sender queries (8 returned content, 16 threads total)
• Items extracted: ~58
• Items surfaced: 8
• Items skipped: 50
• Read time: ~9 min