AI Got Cheaper This Week. Microsoft Banked the Check.

OpenAI cut prices and slashed its error rate this week. Google took the #1 benchmark spot and undercut the old leader on cost. A Wall Street economist checked what any of that has actually done for a company's bottom line — and found nothing, outside the handful of firms selling the infrastructure.

That's the split running through this week's signal. Every model vendor has a "cheaper and faster" story right now, and the stories are true — OpenAI, Google, and the benchmark data all back each other up. But Apollo's own chief economist ran the numbers on the S&P 493 — the index minus the Magnificent Seven — and found revenue-per-worker and profit margins both flat, same week the AutomationBench refresh showed the best agent still finishes barely half its assigned workflows. Meanwhile the checks that are actually clearing are landing somewhere specific: Microsoft just told investors $10 billion of its quarterly revenue came from one customer, and Amazon's paper gain on its Anthropic bet already clears the GDP of a mid-sized country. Cheaper isn't the same as profitable. Here's where this week's money actually went.

Models + launches

OpenAI's GPT-6.1 Terra landed with the kind of numbers that used to take two model generations: API pricing 50% below GPT-5.6 Terra and GPT-6.1 Sol, at $1/$5 per million input/output tokens, while matching GPT-6 Astra's output quality at roughly a tenth of Astra's per-token price. Computer-use tasks run more than 3x faster than both Astra and Sol. The error rate at low reasoning effort dropped from 53.6% to 16.7%; at max effort it fell another 18 points, to 4.3% [1]. These are OpenAI's own internal benchmark numbers, not an independent audit, so read them as the vendor's best case — but even with a discount applied, this is a real jump, not a rounding adjustment.

Google answered within 72 hours. Gemini 4 Carbon took the #1 spot on Artificial Analysis's Intelligence Index v5 with a score of 59, dethroning Claude Opus 5.5, while running about 23% cheaper per task — $7.76 versus Opus 5.5's $10.04 — using 29% fewer output tokens (123,000 versus 173,000) for comparable results, priced at $4/$20 per million input/output tokens with a 50% launch discount still stacked on top [2]. If you picked a model last quarter based on a price-per-task eval, that eval is already stale. I wrote about this exact churn ten days ago when Google's own agents started routing to Claude instead of Gemini — the leaderboard keeps moving, which is the actual argument for not marrying one vendor.

Tooling shifts

Zapier and Fireworks re-ran AutomationBench — 657 tasks across 47 simulated business apps, built from patterns in more than 2 billion monthly tasks across 3.7 million real companies — and the results are the most useful gut-check I've read all week. The best model, Gemini 4 Argon, completes 51.3% of workflows end-to-end. Claude Sonnet 5.5 is second at 44.7%. GPT-6 Sol gets to roughly 80% of flagship GPT-6 Astra's score at about a sixth of the cost, which sounds like a great trade until you read the actual failure mode: "the agent reports success while the data says otherwise." That's the majority of failures in the benchmark, and it's the exact gap between a good demo and something you'd trust with a real customer's order [3].

Vercel had its own two-part cost story. Agent sessions now cost 67% less — live, no migration required — through better prompt caching and more efficient context handling on long threads [4]. Vercel CI, in beta, only re-runs the parts of your pipeline your change actually touched instead of the whole suite, already chewing through close to 8 million tasks in testing at 2–6x the speed [5]. Both are real, both ship today. Put next to AutomationBench's "false confidence" finding, the pattern is consistent: the infrastructure underneath agents is getting cheaper and faster every week. Whether the agent running on top of it actually finished the job correctly is a separate question nobody's benchmark says yes to yet.

SMB angles

Apollo's Torsten Slok published five charts that should be required reading before your next AI budget conversation. Revenue per worker is climbing for the Magnificent Seven and flat for the rest of the S&P 500. Profit margins show the same split. Productivity growth, economy-wide, isn't moving. His conclusion, in his own words: "the technology has delivered zero financial payoff to corporate America outside of tech" [6].

Source: Apollo Global Management — No Signs of an AI Payoff Outside the Tech Sector

Zapier shipped the first practical response I've seen to the obvious follow-up question — if you can't measure the payoff, can you at least measure the spend? Its new AI Limit Bot DMs an employee at 75% of their monthly AI budget asking what they're working on. One sentence of explanation auto-bumps the limit. Two or more bump requests in a month notifies the manager; three straight months of needing one promotes the employee to a higher tier. Nobody gets cut off mid-task. The stat that makes this worth copying: 52% of managers say their org spends over $100k a month on AI tools, and only 34% of individual contributors know where that money is actually going [7]. Before you can answer Apollo's question about payoff, you need Zapier's answer to the simpler one — where did it go. That's the first audit I run in any SMB automation engagement: not "is the AI working," but "can anyone in this company currently answer that question with a number."

Adjacent to watch

If the payoff isn't showing up in the S&P 493's margins, it's worth asking where it is showing up. Microsoft answered that one directly: OpenAI alone generated $10 billion of Microsoft's $92.3 billion in quarterly revenue last quarter — nearly 11% of the total, from one customer, on a pace that annualizes toward $40 billion a year. Outstanding receivables from OpenAI climbed to $7.2 billion, up from $6 billion three months earlier [8].

Amazon's version of the same story is even starker. Its roughly $18 billion bet on Anthropic — split between convertible notes and two rounds of preferred stock — includes convertible notes now carried at $97.9 billion on Amazon's own books as of its last quarterly filing. The catch that makes this a two-way street: Anthropic has committed to spend at least $110 billion on AWS over the next decade [9]. Anthropic owes Amazon more than Amazon's stake is worth being happy about owning. Everyone in this chain is technically winning on paper, and the paper all traces back to one balance sheet.

The models got 50% cheaper this week. Nobody's profit margin got 50% better.

Put this week's efficiency claims next to this week's money flow and the shape gets obvious:

flowchart TD
  Start([This week's signal]) --> Cheaper[Models got cheaper and faster]
  Start --> Paid[Someone got paid more]
  Cheaper --> C1["GPT-6.1 Terra:<br/>50% cheaper, 3x faster,<br/>error rate 53.6%→16.7%"]
  Cheaper --> C2["Gemini 4 Carbon:<br/>new #1 index,<br/>23% cheaper than Opus 5.5"]
  Cheaper --> C3["AutomationBench:<br/>best model still fails<br/>49% of workflows"]
  Paid --> P1["Microsoft: $10B/quarter<br/>from OpenAI alone"]
  Paid --> P2["Amazon: $97.9B notes value<br/>within $18B Anthropic stake"]
  C3 --> Verdict[[Apollo: revenue-per-worker<br/>flat for the S&P 493]]
  P1 --> Verdict
  P2 --> Verdict

None of this means the price cuts are fake or the benchmark wins don't matter — a model that's 50% cheaper and three times faster is a real input cost change for anyone running volume through an API. It means the leap from "the model got better" to "my business got better" still has at least one unverified step in the middle, and this week two separate data sources — one counting workflow completions, one counting S&P 500 margins — landed on the same answer. If you're the one signing the AI budget, Zapier's cost-governance bot is the cheap half of the fix. The expensive half is doing what I do with every SaaS-stack audit: actually tracing the dollar from the subscription to the outcome it was supposed to produce, instead of assuming the vendor's benchmark slide is your P&L.

Sources

[1] OpenAI — Introducing GPT-6.1 Terra — openai.com [2] Artificial Analysis — Gemini 4 Carbon is the new #1 — artificialanalysis.ai [3] Zapier — AutomationBench joins Fireworks' Specialized Intelligence Index — zapier.com [4] Vercel — Vercel Agent sessions now cost 67% less — vercel.com [5] Vercel — Introducing Vercel CI — vercel.com [6] Apollo Global Management (Torsten Slok) — No Signs of an AI Payoff Outside the Tech Sector — apollo.com [7] Zapier — How we built an AI cost-management bot — zapier.com [8] TechCrunch — Microsoft recorded $10 billion in revenue from OpenAI — techcrunch.com [9] TechCrunch — VCs will make billions on Anthropic's IPO, but Amazon is set to make much, much more — techcrunch.com

The short version

• GPT-6.1 Terra: 50% cheaper, 3x faster on computer-use tasks, error rate cut from 53.6% to 16.7% at low reasoning effort
• Gemini 4 Carbon took #1 on Artificial Analysis's Intelligence Index v5, running about 23% cheaper per task than the model it dethroned
• The AutomationBench refresh found the best agent completes only 51.3% of real business workflows — the rest mostly fail by falsely reporting success
• Apollo's chief economist: revenue-per-worker and profit margins are flat across the S&P 493 even as AI spend keeps climbing — "zero financial payoff" outside the tech sector
• Zapier's new AI Limit Bot auto-manages budget overages with a Slack DM instead of a hard cutoff — useful given 52% of managers report $100k+/month AI spend and only 34% of ICs know where it goes
• Microsoft pulled $10 billion — nearly 11% of its quarterly revenue — from OpenAI alone, annualizing toward $40B
• Amazon's convertible notes in Anthropic — part of its ~$18B stake — are already carried at $97.9B on its books, even as Anthropic owes AWS at least $110B over the next decade