The one-screen version: on October 26, OpenAI shipped GPT-6.1 Terra — a mid-tier model built for computer use and agents — at $1 per million input tokens and $5 output, half the price of GPT-6.1 Sol, and by OpenAI's own numbers roughly 9x cheaper per agent task than Claude Opus 5.5. A day earlier, Meta launched a record ~$95B equity raise to keep its data-center build-out funded. Put them side by side and you get the single most useful fact a founder can carry into Q4: **the cost of using frontier agents is collapsing while the cost of building them is going vertical.** Both facts favor the team of one — for opposite reasons. Here's each story, and the one move each one asks of you.

1. OpenAI–Terra: the price of an agent run just got cut in half#

GPT-6.1 Terra is OpenAI's cheaper, computer-use-tuned model, and the headline is the sticker. API pricing is $1/M input, $0.05/M cached input, $5/M output — 50% below both GPT-6.1 Sol ($2/$10) and GPT-5.6 Terra ($2/$12). It's live now in ChatGPT Work and Codex, and developers call it as gpt-6.1-terra.

The capability story is the one to read carefully, because it's where the marketing and the measurement part ways. OpenAI's own charts put Terra at 60.0% on Agents' Last Exam at max effort, for about $1.01 per task — against Claude Opus 5.5's 63.2% at about $8.82 a task. That's the ~9x-cheaper claim: a ~3-point accuracy gap for roughly an order of magnitude less money. On OSWorld 2.1 (offline subset), OpenAI lists 66.7% at medium effort and 74.2% at max. These are OpenAI's self-reported numbers on OpenAI's chosen benchmarks, and cross-vendor scores never line up cleanly — treat them as directional, not a verdict.

What it changes for you: the "always reach for the smartest model" default is now a line item worth re-examining — for agent and computer-use workloads specifically, where you pay per step and the steps multiply. If your agent fires thousands of tool calls a day, a 9x cost cut on a 3-point accuracy give-back may be the best margin trade on your P&L this quarter. The discipline is unchanged: don't take the launch chart's word for it, run the two models head-to-head on your eval harness and your token mix. We did the decision tree on exactly this tradeoff — GPT-6.1 Terra vs Claude Opus 5.5 for computer use and agents — and the short version is: route by task, don't pick a side.

2. Meta–$95B: the other half of the picture#

The same week the usage price fell, the build price got a number attached to it. Meta launched a plan to raise up to ~$95 billion in equity for AI infrastructure. The structure is the interesting part: a first ~$35B tranche (about $22.5B in Class A common via forward agreements plus ~$12.5B in mandatory convertibles, run by Morgan Stanley and Goldman Sachs), then a **~$60B at-the-market program in new nonvoting shares — which lets Meta pull in tens of billions without diluting Zuckerberg's voting control. It's reportedly Meta's first stock issuance since 2013, and it comes against guided 2026 capex of $135–145B** and a negative Q3 free cash flow (~-$2.1B) despite ~$82B in cash.

Then the detail that should make you pause: Nvidia, AMD and Broadcom have signalled nonbinding interest in buying up to ~$5B of the shares combined. The chip vendors are helping finance one of their largest customers — the same circular-financing pattern that's been quietly knitting the AI capital stack together all year, the way Anthropic's own disclosed economics hinted at.

What it changes for you: nothing you do — and that's the point. When a hyperscaler has to issue stock for the first time in thirteen years to afford the next data center, the capital intensity of owning the model layer has moved permanently out of reach for a bootstrapper. That isn't a defeat; it's a clarifying constraint. Every dollar Meta, Nvidia and the rest pour into the floor is a dollar that lowers your cost to rent capability on top of it — Terra's price cut is literally funded by this arms race. Rent the capability, never the capex. The one thing to watch is concentration risk: when your inference bill rides on a market this circular, keep a second provider warm so a single wobble isn't your outage.

The coda: the price war is global#

One more data point so you don't read Terra as a purely American story. In Beijing, Approaching.AI raised over RMB 1.3B (~$195M) in Series B — its stated purpose is lowering inference costs. The same force pushing OpenAI's prices down is capitalized on three continents at once. For a founder shipping to a global audience (and if you check your referrers, Chinese answer engines are already sending readers), this is the tailwind of the decade: the raw cost of intelligence is falling in every currency, funded by people who are not you.

The one line to take into the week: capability is getting cheaper by the month and infrastructure is getting more expensive by the quarter — so build on the layer, never as the layer. Price your agents against Terra this week, keep a second model warm, and let the giants fight over the capex. Last edition is here; if you're choosing the framework to run these agents in, here's what actually changed this fall.