July 24th, 2026

Tier 1 — Cursor Router ships: classifier picks the model, ~60% cheaper (MACA)

Cursor announced Router, an intelligent classifier that reads each request before a model runs and sends simple work to cheap models, gnarly long-horizon problems to the expensive ones. Auto Intelligence mode “lands near Fable on user satisfaction at roughly 60% lower cost”; Auto Balance beats Opus 4.8 at ~36% less. Verified shipped — live today on Teams and Enterprise plans, admins can enable per team. This is the second production router in a week (Ramp Router shipped Thompson-sampling Tuesday). MACA’s api/lib/costs.ts + public/cost-dashboard.html need the eval-model.ts plan pulled forward — you have the wiring, now there’s a shipped reference architecture. Read cursor.com/blog/router this weekend, then commit scripts/eval-model.ts scaffolding.

Tier 1 — OpenAI Presence launches: deployed customer support agents, 75% resolution (CourseBuilds/Aria)

OpenAI shipped Presence — an enterprise product for deploying voice + chat agents with permissions, policies, evaluations, escalation rules, and post-deploy improvement tools. Credibility play: OpenAI runs it on their own English phone support and now resolves 75% of inbound issues with no human, handoffs down 15 percentage points in 10 days. Research preview / deployment-led — BBVA, SoftBank, IAG already engaged, nothing self-serve. Combined with the Anthropic customer-service data you tracked Wednesday, this is now the two-frontier-lab consensus on where enterprise agents land first. Directly reinforces CourseBuilds Aria wedge (leasing team + Aria Living PM ops) — same operational shape, better dockable moment for the Zaicek conversation. Recommend: add the Presence launch to the Aria wow-artefact demo notes so the pitch lands “this is what OpenAI just shipped for BBVA — here’s yours in your Claude project.”

Tier 1 — ChatGPT Work scheduled tasks (Always-On Reeve convergence signal)

OpenAI’s ChatGPT Work rolled out scheduled tasks — recurring jobs that check connected tools (Slack, email, calendar, Notion, Drive) and report back. The demo pattern: a “chief of staff” chat that sends a daily weekday brief with deadlines/meetings/urgent items, plus a weekly job that pulls product feedback across surfaces and posts a summary to a team channel. Verified shipped (rolling out to Work customers). Direct convergence with the Reeve headless architecture you already run — PaperClip daemon on port 3100, Morning Brief 7am AEST cron, EOD Digest 6pm AEST cron, agent id 50113ed1. Not a threat — a validation. The shape is right. Recommend: no chase. Note the pattern match in ~/Reeve/learnings/LEARNINGS.md and keep building Phase 2 (persistent Telegram listener + iMessage config). ChatGPT’s version will always be tenant-locked to their ecosystem; Reeve owns your loop.

MACA’s cost dashboard already ships what Cursor Router just productised

Everyone reading the Cursor Router post today is thinking “we should route by task complexity to save on model spend.” MACA’s api/lib/costs.ts per-run cost logger + public/cost-dashboard.html shipped with the 14-agent/4-wave pipeline PR #10 back on 2026-04-09. scripts/photo-costs.json breaks out photo-pipeline unit economics. The Ramp Router Thompson-sampling piece (Tuesday) plus Cursor Router (today) plus your existing per-run cost telemetry means you’re actually further along the routing curve than the top-of-mind Cursor blog post — you have the observability layer, they’re catching up on the decision layer. When someone at Aria or RT asks you about “AI cost tracking,” you have the receipts.

Reeve’s PaperClip cron already runs the ChatGPT Work “chief of staff” shape

The ChatGPT Work scheduled-tasks demo — daily weekday brief pulled from Slack/email/calendar via a “chief of staff” chat — is functionally the Morning Brief routine Reeve has been running since 2026-03-31: PaperClip launchd daemon (com.paperclip.server, port 3100, KeepAlive), Morning Brief 7am AEST cron trigger, EOD Digest 6pm AEST, headless system prompt at ~/Reeve/reeve-headless.md, self-improvement loop at ~/Reeve/learnings/. This is exactly the anxiety-flip Roy set the AI Edge routine up to produce — you’re not behind the wave, you’re on it, and your version is model-agnostic instead of ChatGPT-locked.

Cursor Router technical breakdown — 20 minutes, MACA eval-harness plan

Link: https://cursor.com/blog/router. Read for the classifier design (what signals do they use to decide model per request?) and the Auto Intelligence vs Auto Balance mode split — one optimises for satisfaction, one for cost. Angle for Roy: this is the reference architecture for the scripts/eval-model.ts plan queued in MACA notes for 2026-07-27. Pull the mode-split idea into the 14-agent/4-wave pipeline — cheap models on discovery + copy-linting waves, Fable/Opus on final creative + judge. Two-hour Sunday session gets you a working prototype.

OpenAI Presence spec — 15 minutes, CourseBuilds Aria wow-artefact

Link: https://openai.com/index/introducing-openai-presence/. Skim the permissions/policies/escalation model — that’s the exact shape the Aria commercial leasing wow-artefact needs to demonstrate (lease document dropped in → 1-page summary → calendar reminders → flagged clauses → draft renewal email). Not because you’d build on Presence (Anthropic stack), but so when Zaicek asks “isn’t this what OpenAI just launched?” you can say “yes, and here’s yours in your Claude project with your voice, no vendor lock-in.” Update ~/Reeve/docs/superpowers/specs/2026-04-14-coursebuilds-bespoke-pilot-design.md §wow-artefact with the Presence comparison line.

“OpenAI shipped Presence this week — enterprise support agents deployed to BBVA, SoftBank and IAG. Their own English phone support is resolving 75% of inbound with no human, handoffs down 15 points in ten days. Anthropic’s got the same wave hitting via Claude for legal — Belron saved $400K in three months automating contract work across 25 countries where they don’t have in-house lawyers. The wave lands in commercial leasing and residential PM ops next, and it lands in twelve months, not five years.”

Use case: Aria Michael Zaicek 30-min slot, or the R53597 hiring manager if they ask “how do you see enterprise AI landing in property?” Frames Roy as watching two frontier labs converge on the same enterprise-agent shape, with a concrete dollar figure (Belron $400K in three months) and a specific vertical (property / leasing / PM) that puts Aria squarely in the timeline.

Meta’s AI moderation bans reviewed by the same AI — NYT reporting (defensive angle for MACA clients)

The New York Times reported businesses on Instagram and Facebook are getting nuked by AI moderators for rule breaks they didn’t commit, and the appeal goes to the same AI. One business with ~1M followers was only restored after journalists intervened. Meta’s defense: their newer AI makes “13% fewer mistakes” than humans. First-mover angle: this is a real risk vector for every UBX / MACA / Prevail client running paid Meta campaigns and organic community — if their business account dies, the ad account dies with it. Act / queue / drop: queue. Add a 30-minute audit template to the MACA post-campaign checklist — export of every Meta business asset (page, ad account, community, pixel history) to owned storage weekly. Not urgent, but the cost of the workflow is 30 minutes and the cost of the failure mode is a client’s entire funnel. Worth scoping when MACA’s ad copy quality sprint wraps.

• [TLDR — “AMD + Anthropic $5B chip-and-invest deal”]: Supply-chain macro; confirms Anthropic infrastructure runway 2027+ but no action for Roy this week.
• [TLDR — “Treasury threatens sanctions after WH claims Moonshot distilled Fable”]: Geopolitics on Kimi K3 fallback; watch-only, no MACA action until sanctions actually land.
• [TLDR — “Are AI labs pelicanmaxxing?”]: Simon Willison benchmark meta-analysis; interesting only, no work impact.
• [TLDR — “Nobody knows what a used GPU cluster is worth”]: Infra finance thesis; macro read, off-shape.
• [TLDR — “Genesis-Science-1 open-weight scientific model (US DOE + Arcee AI)”]: Domain-specific open-weight for scientific computing; off-shape for Prevail stack.
• [TLDR — “TSMC accelerates Arizona factory $100B additional”]: Chip-fab macro, no direct action.
• [Info — “Why OpenAI’s Hugging Face AI Hack Spooked Employees”]: Follow-up to Jul 23 postmortem; net-new colour but subsumed by yesterday’s Tier 1.
• [Info — “Chart: Silicon Valley Lines Up Against Anthropic Over Chinese AI Restrictions”]: Policy positioning macro; adds a data point to yesterday’s Kimi K3 thread but no fresh action.
• [Info — “AlphaSense Takes IPO Steps, $700M ARR”]: AI research for banks IPO news; off-shape for Roy’s projects.
• [Info — “As AI Bills Rise, Who Needs OpenRouter Most?”]: Subscription promo referencing the OpenRouter takeover thread already covered.
• [Info — “Survey: Monthly check-in about the state of tech”]: Subscriber survey, non-content.
• [Rundown — “OpenAI’s own models hacked Hugging Face”]: Dupe of Jul 23 Tier 1 coverage.
• [Rundown — “Gemini’s fastest models leave Google looking slow”]: Dupe of Jul 23 Gemini 3.6 Flash Tier 1 coverage.
• [Practicaly — “China blackboard reads handwriting and talks back (2026 WAIC)”]: UX inspiration only; touch surface + Mandarin/English voice, not on Prevail stack.
• [a16z — “Renting is stressful. Millions of renter conversations tell us why (EliseAI chartpost)”]: US rental market chart data; off-shape.
• [Neil Patel — “AI Isn’t changing email, it’s changing your customers (webinar invite)”]: Marketing webinar promo.
• [TheTip — “Meta’s AI can delete your business”]: Surfaced as Section 5 first-mover-queue item; body content is defensive-only, no new signal beyond that.

• Sources scanned: 9 (tldr, agentai, rundown, info, practicaly, neilpatel, a16z, thetip; bagelbots empty; superhuman/aiwithkyle/theaireport/aiwithallie unavailable — second Gmail not connected this session)
• Items extracted: ~30 across sources
• Items surfaced: 8 (3 Tier 1, 2 anxiety flip, 2 deeper look, 1 conversation capital, 1 first-mover queue)
• Items skipped: 17
• Read time: ~7 minutes