📅 October 27, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 32
You're in the loop — the price of frontier AI keeps falling off a cliff: OpenAI's newest model matches last month's flagship at roughly a tenth of the cost. And in a twist that shows the other side of the ledger, Anthropic finally filed to go public — an S-1 that puts its record losses on the record for the first time.
Today: a cheaper-faster model, Meta's record raise and Microsoft's buyout wave, why nations are being forced to pick an AI bloc, and a Workbench on routing every task to the right-priced model.
🔁 The Loop
Cheaper AI, a record raise, and a buyout wave
An original InTheLoop illustration: frontier capability is getting faster and cheaper in the same breath.
OpenAI's GPT-6.1 Terra makes frontier-grade AI cheap. OpenAI launched GPT-6.1 Terra, a mid-tier model priced at $1 per million input tokens and $5 output — 50% below GPT-6.1 Sol — that runs tasks more than 3x faster and matches flagship GPT-6 Astra on Agents' Last Exam at about a tenth of the per-token price. It reaches ChatGPT and the API on October 27. Expect rivals to answer within weeks as the price floor keeps dropping. See the benchmarks.
Meta asks Wall Street for a record $95 billion. Meta kicked off its largest-ever equity raise — roughly $35 billion in common and preferred stock now, plus an at-the-market program of up to $60 billion — to bankroll AI data centers after committing about $825 billion in future spending and posting its first negative free cash flow in over a decade (−$2.12B last quarter). Shares slipped about 1.4%. Read the details.
Microsoft dangles buyouts to 60,000 staff. Microsoft offered voluntary buyouts to about 60,000 employees — roughly 27% of its workforce, VP-level and below — across US, UK, Canada, and India teams, with US packages starting at eight weeks' pay. Staff must decide in December and exit by the 31st. The offer lands as Microsoft pours a projected $145 billion into AI this fiscal year. Here's the memo.
Worth your time: why Meta's first negative free-cash-flow quarter in a decade has analysts bracing for even more AI raises.
🌊 Deep Current
The choice no country wants to make
An original InTheLoop illustration: a world increasingly asked to stand on one side of the AI divide.
Two clubs, one planet. The AI cold war now has membership cards. The US-led Pax Silica and China's newly launched WAICO are both signing up countries — WAICO opened with 29 founding members including Pakistan and Russia, with Indonesia since joining. Beijing's reach is already wide: China has funded AI projects in at least 77 low- and middle-income countries, per AidData.
Why it matters. This isn't a treaty — it's about who supplies your chips, your models, and your standards. Washington has made the stakes explicit: the State Department says a country that substantively engages with a rival "vision for superintelligence" simply can't be trusted as a Pax Silica partner. Pick wrong and you may lose access to the frontier.
Kazakhstan's AI minister put the squeeze plainly: "We don't want this binary choice."
The awkward middle. So most countries are trying to live in both worlds. Kazakhstan joined both blocs within three weeks while holding a $10B Nvidia deal and record Chinese investment; Saudi Arabia built a platform on a Chinese open model yet also signed with Nvidia; Japan leans on Washington for security but quietly runs cheaper Chinese models for low-risk work.
The bottom line. For businesses, the bloc map is becoming a procurement map: your model vendor, cloud, and chip supply may soon carry a geopolitical label whether you chose one or not. Watch the fence-sitters as export rules harden — the "part of both" era may have a short shelf life.
🛠️ The Workbench
Route every task to the cheapest model that can nail it
New tiers like GPT-6.1 Terra mean you rarely need your priciest model for routine work. Here's a five-minute routing habit that trims your AI bill without dropping quality.
List your recurring task types — drafting, summarizing, coding, research — and the real quality bar each one needs.
Assign a default tier to each: a cheap-fast model (say, GPT-6.1 Terra at $1/$5) for drafts and summaries; a flagship only for complex reasoning or production code.
Write a one-line "promotion rule": escalate to the flagship only when the cheap model fails a specific check — facts wrong, tests red, logic off.
Bake the model name into your prompt templates so the choice is automatic, not decided ad hoc each time.
Review spend weekly. If the cheap tier is handling 80% of your volume, you're routing right.
Sample Prompt: "You are my model-router. Given this task: [paste task], recommend the cheapest capable tier (fast / standard / flagship), name the one check that would justify escalating, and estimate relative cost versus the flagship."
🗣️ Overheard
What the timeline's buzzing about
📱 Siri grows up: Apple shipped iOS 27.2 with its Siri AI assistant in five new languages — French, Japanese, Korean, Portuguese, and Spanish.
💸 The $29 ceiling: Accenture capped staff Claude spending at $29 a month after some power users blew past $1,000.
🧠 Kimi climbs: Moonshot's new Kimi K3.1 Pro jumped to 55 on the Artificial Analysis Intelligence Index, landing 4th — just behind Claude Opus 5.5.
🤖 Learned by watching: Alibaba's Amap open-sourced λ0, a whole-body humanoid model trained on 500 hours of human video that hit 62.5% on real-world loco-manipulation.
🔓 Zombie agents: Security researchers found 76% of enabled AI agents hadn't run in 60 days yet still held live credentials.
🔎 Fresh Finds
Five tools worth a look
🎬 Jockey by TwelveLabs: a conversational agent that searches, clips, and assembles across thousands of hours of your video.
🔌 Rube by Composio: a universal MCP server that lets your AI actually use 500+ apps straight from chat.
📧 Inbox Zero: an open-source AI assistant that triages your Gmail and drafts replies for you.
🍰 Pancake: lets AI agents run B2B go-to-market outreach through a single MCP connection.
📰 daily.dev for agents: a discovery feed where your AI agents find what to read and explore next.
★ = sponsored placement, if any.
🧪 Prompt Lab
The model-router prompt
Paste this before any task to decide, honestly, whether you actually need your flagship model — or whether a cheap-fast tier will do the job for a fraction of the cost.
You are my AI model-router. I'll describe a task; you decide the cheapest tier that can do it well. TASK: [paste your task here] QUALITY BAR: [e.g. "must be factually airtight" / "rough draft is fine"] Return: 1. Recommended tier: fast / standard / flagship — and why in one line. 2. The single check that, if failed, should trigger escalation to the next tier up. 3. A rough cost multiple vs. running this on the flagship (e.g. ~0.1x). 4. One risk of going too cheap for this specific task.
Want a banner to match the theme? Try this in a "paper-cut diorama" art style:
A paper-cut diorama, layered cardstock with soft shadows: a tiny robot courier sprinting along a winding ribbon road carrying a glowing parcel, past small signposts, deep-indigo and warm-amber palette, generous negative space, no text.
⏪ Rewind
Yesterday’s most-opened link
Readers couldn't stop clicking the rundown of how far frontier-model prices have fallen this year — and GPT-6.1 Terra just pushed the floor lower again. Revisit it.
Stay in the loop — the InTheLoop team