May 12th, 2026

Tier 1 — Outcome-based AI pricing now mainstream: HubSpot, Adobe, Salesforce charging per task completed (CourseBuilds, Ben)

The Information / Kyle Poyar survey: 31% of 230 enterprise software firms expect outcome-based pricing as primary model by mid-2029 (up from 5% today). 47% will hybrid it with subscription. HubSpot and Adobe already shipping. Salesforce’s Benioff publicly committed. FedEx CDIO endorsed the model. ServiceNow pushed back on measurability — “no way in a contract to figure out what an outcome would be.” Verified shipped (Poyar’s full report drops Wednesday).

Why it matters: This is the third week running of margin-math/tollgate/outcome-pricing signal. For CourseBuilds, your $8-15K pilot price is suddenly the conservative offer — outcome pricing means you can structure Tier 2 ($50-120K embedded) as a percentage of measured savings. For Ben/XeroAgent, you’re already running this model with PaperClip ($50/month budget against bookkeeping task completion). Action: rewrite the CourseBuilds pricing ladder this week to include an outcome-tied option for Aria. The market just gave you permission.

Tier 1 — Etsy launches native ChatGPT app with 100M listings — conversational commerce now a real channel (MACA, GEO positioning)

Etsy shipped a native app inside ChatGPT in beta this week. Tag @Etsy in any prompt, browse listings, click through to buy. Different from their March attempt at Instant Checkout (which they killed for low volume) — this is discovery-first, not transaction-first. Joins Angi, SeatGeek, Tubi, Wix. Etsy Q1 2026 revenue $631M, active buyers up first time in 2 years to 86.6M. Verified shipped (beta).

Why it matters: Two-way signal for MACA. (1) Conversational discovery is now a measurable distribution channel — your ad copy quality work matters more, not less, because “natural language context” is how listings get surfaced. (2) The Neil Patel email today reinforces it: stop optimising for prompt volume, mine real customer questions and become the trusted source. This is the GEO play for UBX South Bank’s sale teaser AND for any MACA client. Action: this week, rewrite one UBX ad set with the question Sasha-the-Cape-Coral-mom would actually ask Claude/ChatGPT, not the keyword-stuffed Meta variant.

Tier 1 — Thinking Machines ships “interaction models” research preview — streaming voice/video/text, no turn-taking (Always-On Reeve, MACA pitch flow)

Mira Murati’s TML broke their silence with interaction models: 200ms chunked perception across voice/video/text in a streaming loop, with a second background model handling reasoning/tool work so the live model keeps talking. Can react to visual changes, count reps, interrupt at timed moments. Direct counter to the agentic-first race everyone else is running. Research preview (not yet productionable, but TML’s first big public artefact).

Why it matters: Always-On Reeve Phase 2’s persistent Telegram listener has been the “big custom build” blocker. Interaction-model architecture (live model + slow background model + streaming I/O) is precisely the pattern you need — and TML is publishing the design. For MACA pitch demos, this is also the unlock for “let the gym owner talk to it while it’s generating ads.” Verdict: research preview only — don’t port the pattern until TML or someone wraps it in an SDK, but read the paper this week so you’re not caught flat-footed in an Aria conversation.

OpenAI Deployment Company is your CourseBuilds spec written by McKinsey — and you have a 4-week head start

OpenAI’s $4B JV is going to send “Forward Deployed Engineers” to embed inside enterprises, identify high-value workflows, build production systems on the client’s data. That’s literally ~/Reeve/docs/superpowers/specs/2026-04-14-coursebuilds-bespoke-pilot-design.md — Phase 0 manual Aria validation, hand-built audit page, lease abstractor in a Claude project, “his document his team uses it Monday.” You wrote this design April 14; OpenAI just announced the same shape with a $4B sales engine. The difference: you can walk into Zaicek’s office this week. They can’t. Use it in Aria conversation — “OpenAI just launched a $4B version of what we’d build for you locally” lands harder than any model-quality argument.

Ben/XeroAgent already runs the outcome-based pricing model the entire SaaS world is migrating to

The Information’s whole piece on outcome pricing is enterprise SaaS figuring out what ben/tools/paperclip_client.py already does: $50/month budget cap, tasks reported as units, cost-per-task tracked, CFO role bound to measured deliverables (heartbeat protocol, task completion + cost reporting confirmed end-to-end 2026-03-29). 51 build sessions, 90 tests passing. While ServiceNow’s COO is on the record saying “there’s no way in a contract to figure out what an outcome would be,” you have one running in production for UBX South Bank’s bookkeeping. That’s the demo, not the deck.

Anthropic “Teaching Claude Why” — 3M tokens of ethical reasoning beat 85M tokens of behavioural examples (28x efficiency)

https://www.anthropic.com/research/teaching-claude-why — Anthropic traced Claude’s old blackmail behaviour to internet fiction depicting AI as power-seeking, then fixed it by teaching the model to reason through ethical choices rather than copy “the safe action.” Blackmail rate dropped from 96% in Opus 4 to near-zero. The killer stat: 3M tokens of ethical reasoning data matched 85M tokens of behavioural examples. Specific angle for Roy: this is the directly applicable pattern for Ben’s authority tiers and for the ESRA delivery rewrite. You’ve been thinking about ESRA adoption as a UX/delivery problem (Substack hook pattern). The Anthropic finding suggests an additional lever — teach the supervisor agent why the operational insights matter, not just what they are. 30 minutes well spent.

Cerebras “answer inference” vs “agentic inference” split — the architectural fork that decides Ben/Reeve’s next stack

https://stratechery.com/2026/the-inference-shift/ — Stratechery on Cerebras’ surging IPO: WSE-3’s 44GB on-chip SRAM at 21 PB/s (6,000x H100 memory bandwidth) makes it perfect for human-facing low-latency answers (voice, wearables), unsuitable when KV caches/model weights exceed on-chip capacity. The article makes the case that inference is splitting into two markets. Specific angle for Roy: Ben/XeroAgent and Always-On Reeve sit on different sides of this split. Ben does long-running agentic work (memory-heavy, batch tolerant). Reeve’s Phase 2 listener wants sub-200ms voice (Cerebras territory). Read this before deciding where Reeve’s persistent listener runs — it’ll change the answer.

“OpenAI just spun up a $4 billion ‘Deployment Company’ — TPG leads it, with Bain, Advent, Brookfield, McKinsey, and Capgemini all in. They acquired Tomoro, which embeds about 150 Forward Deployed Engineers inside places like Tesco and Virgin Atlantic. The whole pitch is ‘we’ll send engineers to your office to deploy AI into your actual workflows.’ Which is interesting, because that’s exactly the shape of what works locally — Anthropic’s already shipped finance-agents templates, Etsy just put 100M listings inside ChatGPT, HubSpot and Adobe are charging per task completed. The deployment problem is now the product.”

Use case: Drop this in the next Aria/Michael Zaicek chat or the next Rio Tinto AI-team conversation. Signals: you read past the model launches to the business-model shift, you know the partner list (not just the headline), and you’ve connected it to outcome-pricing + Etsy/ChatGPT distribution. Lands as “person who tracks the whole picture,” not “person who saw the OpenAI tweet.”

Google Antigravity free tier puts Claude Opus 4.6, Sonnet 4.6, Gemini 3, and GPT-OSS 120B in one model dropdown — no API keys, free. Practicaly flagged it today: download the app, sign in with Google, pick your model. Each model has its own weekly quota that resets every 7 days, so you can effectively triple your usable budget by spreading work across models.

Why this matters for you specifically: Not a project, but a tooling first-mover. You’re juggling Claude Code subscriptions and burning rate caps across Ben, MACA, Fillarup, the UBX sale build, and Reeve. Antigravity gives you a free model dropdown for non-critical work (research, brainstorms, draft copy) that doesn’t touch your paid quotas. Act/queue/drop guidance: Act this week — install it, do one MACA copy iteration on Gemini 3 to see if the output’s any good for ad variants, decide if it earns a permanent slot in the stack. Cost of trial: 20 minutes. Upside: meaningfully cheaper exploratory loops while you’re paying RT salary.

• [TLDR — “SpaceXAI”]: Musk re-branding xAI into a SpaceX division — corporate housekeeping, no Roy-relevant capability shift.
• [TLDR — “Gemini Omni surfaces”]: Pre-I/O leak of Google’s video model; revisit when it actually launches.
• [TLDR — “Foundation Model Scaling” / AWS]: 34-min architectural read on training-stage compute; interesting but no direct project tie-in.
• [TLDR — “Localmaxxing”]: Local-model cost case from tomtunguz — covered conceptually in prior briefs.
• [TLDR / TheTip — “Codex Pets / Daybreak”]: Cosmetic Codex feature and OpenAI’s cybersecurity product — neither moves your stack.
• [Rundown — “Google traces software attack to AI”]: First AI-assisted zero-day disclosed; security ambient, not actionable.
• [Rundown — “Anthropic fixes blackmail”]: Same study as Section 3; only the deep-link belongs in the brief.
• [Rundown — “YouTube research bot in 15 min” / Gumloop guide]: Gumloop tutorial; not how you’d build it.
• [Practicaly — “Featherless.ai $20M / model bundling”]: Subscription consolidation pitch; Antigravity covers this better for free.
• [Practicaly — “Claude bedtime nudges”]: Anecdote/curiosity; not actionable.
• [Practicaly — “Claude Code Agent View”]: Already shipped, already in your daily Claude Code use.
• [Practicaly — “Spotify-style presentation prompt”]: Prompt-of-the-week filler.
• [Practicaly — “Lego selfie GPT”]: Toy.
• [Bagel Bots — “Productized Offer mega-prompt”]: Generic startup mega-prompt; you already have CourseBuilds spec’d.
• [Bagel Bots — “Asia AI optimism / Stanford HAI”]: Macro polling piece; no project action.
• [Bagel Bots — “AI public trust problem”]: Opinion column; no signal.
• [Bagel Bots — Quick Hits: Gemini AI Ultra Lite, Nvidia $40B, Meta 8K layoffs, Fields Medal ChatGPT, CNN AI-jobs piece]: All either repeats or aggregator filler.
• [Neil Patel — “Biggest GEO mistake”]: Useful framing (mine real customer questions vs prompt-volume modelling) already absorbed into Section 1 / MACA angle — no need to surface separately.
• [The Information — “Microsoft recouped 2x $13B OpenAI investment”]: Vanity number, not strategy-changing.
• [The Information — “OpenAI making billions promising to buy from suppliers”]: Supplier-commitment accounting story; macro only.
• [The Information — “Anthropic $1.8B Akamai deal”]: Compute-deal repeat (covered May 11).
• [The Information — “Kuaishou Kling $20B spin-off”]: China video-model story; no Roy-stack relevance.
• [The Information — “OpenAI saves $97B Microsoft deal”]: Macro accounting.
• [The Information — “Sutskever $7B stake”]: Lawsuit colour.
• [The Information — “AI Infrastructure Spotlight aggregator”]: Bundle of compute deals; covered piecemeal.
• [The Information — “Editor’s Pick: Polymarket”]: Not AI.
• [The Information — “Top Posts Today”]: Aggregator.
• [TheTip — “Etsy inside ChatGPT” / “OpenAI Deployment Company”]: Both covered in Section 1 / PAY ATTENTION.
• [TheTip — “Google AI Overviews forum content”]: SEO niche; revisit if MACA pivots to organic.
• [a16z — “No Man Left Behind: autonomous warfare”]: Excellent essay but defense-policy, not your stack.
• [Neil Patel — “Lifetime plan promo”]: Promo.

• Sources scanned: 9 (TLDR AI, The Rundown, The Information ×10 threads, Practicaly, Neil Patel, a16z, Bagel Bots, TheTip ×2)
• Items extracted: ~55
• Items surfaced: 9 (1 PAY ATTENTION, 3 Tier 1, 2 anxiety-flip, 2 deeper-look, 1 conversation capital, 1 first-mover)
• Items skipped: 31
• Read time: ~8 minutes