Sponsored by
Sponsor:
Our Founder Was Early to AI Winners. Now He’s Tracking What Comes Next.
Over the last 18 months, MavSource’s founder tracked and invested around major AI-related names like Micron +100%, Nvidia +74%, Sandisk +130%, Western Digital +74%, TSM +22%, Broadcom +27%, Okta +35% and Lam Research +39% — an average return of approximately +63% across the group.
Now, he’s bringing the sources and ideas behind those insights to you in one daily email digest. MavSource aggregates the most important AI updates from newsletters, podcasts, company news, AI labs, funding rounds, and more — then summarizes what matters in a simple 5-minute brief.
Past performance does not indicate future results. Informational only; not investment advice.
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #142 WEDNESDAY · 8 JULY 2026
THE HOOK
Here’s the sequence, because the order is the story. Washington delays GPT-5.6. Days later — per FT-reported talks — Sam Altman floats handing the US government 5% of OpenAI, roughly $42.6 billion. Nobody demanded it. That’s the tell. A month ago the labs endured the permission layer. This week they started paying for it — because a gate you’ve helped build is a gate your competitors still have to queue at.
AI NEWS ROUNDUP
This week in AI — four stories, one direction
1 Washington’s 30-day pause is nearly ready — and the labs helped write it. The White House is finalising a voluntary framework with OpenAI, Google and Anthropic: up to 30 days of federal review before a frontier release — advisory, not blocking, per officials briefed on the talks. Anthropic has already committed to pre-release government access in writing. So what: the giants can absorb a 30-day wait. A startup’s launch window can’t. Full story →
2 The 5% stake is reported, not signed — the norm moves anyway. Per the FT, OpenAI has held early talks on giving Washington a 5% equity stake, with every leading US lab paying into an Alaska-style public fund. Conceptual, early, and it would likely need Congress. The point isn’t whether it passes. It’s that a frontier lab now treats paying the state as a reasonable cost of shipping. CNN →
3 Fable 5 starts billing this morning — two facts, straight from the source. Anthropic’s redeployment note and billing docs confirm both. One: Fable 5 now runs on usage credits — $10 in, $50 out, per million tokens — and accounts without credits enabled lose Fable 5 access entirely. Two: requests blocked by the new safety classifier are sent to Opus 4.8 instead, with an in-session notice. Why that notice matters: this week’s Tip. Anthropic →
4 Anthropic now out-earns OpenAI — which explains a lot about story two. A $30 billion run rate against roughly $24–25 billion, per this week’s reporting — built on enterprise contracts, not consumer subscriptions. Why it’s in this issue: the lab best able to afford a compliance moat just became the revenue leader. Voluntary rules look different from the front of the queue. Fortune →
OUR ANGLE
🔭 Permission is becoming a moat — and the big labs are buying The spine of it. A 30-day federal review is a rounding error for a lab running $30 billion in revenue and a death sentence for a startup’s launch window. Any rule with a fixed cost favours whoever is biggest. So the rational incumbent move isn’t to fight the gate — it’s to fund it, shape it, and stand politely behind it. What this week showed. Equity offered (reported). A review window negotiated (confirmed talks — and Anthropic’s pre-release access commitment is already published). A seat taken at the UN’s new AI commission (confirmed, first meeting today). Three different kinds of event — one direction of travel. The honest caveat. The framework hasn’t landed. If it arrives with a threshold set high and no review teeth, ‘buying the gate’ becomes this column’s overreach of the year. The bet. By the end of Q4 2026, at least one US frontier lab markets government review as a customer-facing trust signal — ‘federally reviewed before release’ in enterprise sales material. |
THE THREE SPECIALS
Do · Use · Understand
🎯 PROMPT OF THE WEEK
The Downgrade Test
Expensive model usage persists for one reason: nobody ever forces the task to justify the tier. Fable 5 going credits-only this morning makes the audit urgent — but this prompt pays for itself on any stack, any week.
You are my model-routing auditor. Below is a task I currently run on my most expensive AI model, plus one typical output. 1. Strip the task to its real requirements: reasoning depth, context length, tool use, and the cost of a wrong answer. 2. For each requirement, judge whether a mid-tier or budget model meets it — and how one test would prove it either way. 3. Design that test: same task, cheaper model, and the three things to compare across the two outputs. 4. Verdict: KEEP on frontier, MOVE down a tier, or SPLIT — draft on the cheap model, finish on the expensive one. Rule: “the expensive model feels better” is not a requirement. |
Why it works: it separates what the task needs from what you’re used to — the discipline commercial routing layers apply automatically, done once by hand. SPLIT is where the money hides: drafting at $2 per million tokens and polishing at $50 cuts the bill without touching quality where it matters. Works best on: the frontier model itself — nobody knows the cheaper tiers’ blind spots better.
🛠️ TOOL OF THE WEEK GLM-5.2 The fallback nobody can gate: an MIT-licensed coding model whose weights live on your own disk. ★★★★ / 5 This issue is about gates, so the tool is the exit. GLM-5.2 is Z.ai’s open-weight model — 744 billion parameters, a 1-million-token context — at $1.40 in and $4.40 out per million tokens, or free self-hosted. On Z.ai’s published benchmarks it beats GPT-5.5 on several long-horizon coding tests and comes within a point of Opus 4.8 on FrontierSWE; third-party leaderboard runs point the same direction. Vendor-led numbers either way — apply the usual salt. Use if: your fallback plan has to survive any government’s next letter, and your work is text and code. Skip if: you need vision or PDFs — it’s text-only — or Chinese-origin weights won’t clear compliance; self-hosting answers the data question, not the provenance one. Describe it to a colleague: “Most of the frontier at a sixth of the price — and we keep the weights.” Best use case: the model behind last week’s LiteLLM fallback route. GLM-5.2 → |
💡 TIP OF THE WEEK An outage stops you. A downgrade just sends a notice. One under-read line in Anthropic’s redeployment note: requests the new Fable 5 safety classifier blocks are sent to Opus 4.8 instead, with a notification — and Anthropic itself says the stricter classifier flags more benign coding requests than before. Fine in a chat window. But when the request comes from a script, an agent, a pipeline: who reads the notice? Name your fallback in writing. Every workflow that touches a frontier model should state what it falls back to — and who gets told when it happens. Trust metadata, not the model. Asking a chat session “which model are you?” is unreliable — models get this wrong constantly. The model field in the API response is the ground truth. Why it works: notifications are designed for humans; substitutions in automated workflows only surface in metadata your monitoring actually checks. Where it doesn’t apply: hands-on chat use — the notice does its job there. Glance and move on. Pro move: log the model id on every API call and alert when it changes without a deploy. |
YOUR MOVE
One action. Everything else is optional.
This week in three lines:
Washington’s 30-day review framework could land any day — and OpenAI’s reported response was to offer equity, not objections.
Fable 5 bills credits from this morning — no credits, no Fable. And classifier reroutes to Opus 4.8 announce themselves in a notice your scripts don’t read.
Fixed-cost rules favour the biggest labs — which is exactly why the biggest labs have stopped fighting them.
Your one action: run the Downgrade Test on your single most expensive AI workflow, and reply with the verdict — KEEP, MOVE or SPLIT. I read every response.
(Then, if the “why” is nagging at you: the deep dive makes the full case for what a regulator-shareholder does to this market.)
(Optional, five more minutes: pull your last ten API responses and read the model field. If you can’t say which model answered, that’s your finding.)
—
R. Lauritsen
EDITOR · IPROMPT
P.S. If your Downgrade Test verdict surprises you, that’s the reply I most want to read — the three best verdicts get anonymised and printed next week.
Forward iPrompt → For the colleague who thinks “voluntary framework” means optional. |
See what enterprise-ready AI support looks like
How are leading teams getting AI support to work? We're breaking down the playbook, live July 9.