Source: raw/Why_Your_Company_Doesn_t_Need_the_Most_Expensive_AI_Model.md (AI News & Strategy Daily, Nate B Jones, https://www.youtube.com/watch?v=JlHqTyW8rtA) · raw/AI_Just_Got_Cheaper._Here_s_Why_That_s_a_Problem_for_NVIDIA..md (same channel, https://www.youtube.com/watch?v=1aXBg-qQzgA) · Ramp AI Index, November 2026 update, fetched 2026-11-17 (ai-research/ramp-ai-index-november-2026-new-fronts-2026-11-17.md, https://ramp.com/data/ai-index-nov-2026)

Creator: Nate B Jones (two solo analysis episodes, fetched 2026-11-17); Ara Kharazian, Ramp’s lead economist (first-party data) · Platform: YouTube, Ramp Economics Lab

Ramp’s November index shows business AI spending up 1.65% in October while the average price per million tokens fell 21% and token volume kept rising. Nate B Jones explains the gap with “task saturation”: for much ordinary work several models are already good enough, so buyers test down to the cheapest one that passes. This article records Ramp’s figures as published, Jones’s argument, and his checklist for buying models on evals. The Ramp numbers are first-party; the rest is one analyst’s reading.

Key Takeaways

  • Spend is nearly flat, usage is not. Ramp: AI spending rose 1.65% from September to October, a new peak; token volume growth “continues to outpace token spend as AI gets cheaper and cheaper”.
  • Prices are falling fast. Ramp: the average price per million tokens fell 21% in a month.
  • Task saturation, defined. Jones: “several models can already do a particular job well enough that buying more intelligence at a higher price adds very little or nothing to the results.”
  • Buyers have stopped asking which model. Jones says enterprise customers now pick a cheap model, run evals, and “iterate toward the cheapest available model that passes”.
  • Evals give procurement leverage. If another model passes your tests, you have evidence it is a substitute, and you can negotiate a discount without switching vendor.
  • The premium tier still holds most of the dollars. Ramp: Opus, Fable, Sol and Astra took 69.5% of spend in October, up from 67.7%. Share of dollars is not share of work.
  • New spend is coming from support, which suits cheap models. Ramp: customer success and support went from 0.8% of classified spend in June to 13.3% in October.
  • Your customers will do the same to you. If you sell an AI-assisted service, expect buyers to ask for the saving.

What Ramp published (November 11)

Ramp tracks card and bill-pay spend at its business customers, so the sample is Ramp’s customer base, not the economy.

MeasureOctober 2026Change
AI spending, month on monthnew peak+1.65%
Businesses paying for AI56.8%+1.0 point
Average price per million tokens-21% in a month
Token spend, OpenAI+2.1%
Token spend, Anthropic+0.5%
Median firm, per employee per month$16.32+$0.93
Top 10% of firms, per employee per month$801roughly flat
Top 1% of firms, per employee per month$10.2K+9%
Engineering share of classified spend72%down from summer highs
Customer success and support share13.3%from 0.8% in June
Frontier models’ share of spend69.5%from 67.7%
Paying Anthropic / OpenAI / SpaceXAI44.8% / 40.6% / 6.3%+1.1 / flat / +2.2 points
  • Kharazian’s four standing risks: 1% of businesses drive 80% of revenue to OpenAI and Anthropic; spend is not spreading beyond tech, finance and technical services; prices keep falling; engineering spend “stutters along, despite rising usage”.
  • His read on support spend: these use cases “are a good candidate to rely on cheaper and faster models, rather than the highest-margin and highest-performance models”.
  • His read on competition: SpaceXAI’s adoption gain will turn “this two-way race” into a three-way race “that puts greater price pressure on OpenAI and Anthropic”.
  • Department-level spend is a new dataset in this edition.

Jones relays the same figures. He calls spending “essentially flat for three months running”, citing Ramp’s October report of almost no growth across August and September. He also notes Ramp’s monthly numbers “bounce around a lot”, with some months up 30-40%.

Jones’s argument

  • The arithmetic. A business using 1 billion tokens at an average 5,000. If it moves routine work to cheaper models and its average falls to 5,000. The extra 4 billion tokens produce business results and no supplier revenue.
  • Trading down does not require changing vendor. A customer can move to the same lab’s cheaper model; keeping the customer is not keeping the revenue per token.
  • Contracts are being negotiated from the buyer’s side. Ramp’s October report, as Jones relays it, described customers negotiating enterprise contracts “from very favorable positions”.
  • Where frontier models still earn their price: work the cheaper model fails, and new capabilities. He also notes that iterating two or three times on a cheap model can still cost less than one pass on the expensive one.
  • Lower prices open new work. A task too expensive at a dollar may be worth doing at 10 cents, which he calls the lesson of Jev.
  • Routing is now ordinary. In support, a capable model takes the hard cases and a cheap one “handles almost all of the volume”; “routers are now commonly available”. Google’s Antigravity Orchestrator switches between Gemini 4 Carbon and Hydrogen within a task and, in Google’s own testing, got close to Carbon’s success rate with 24% less quota and 19% less time. He says people using open-weight models report 80-90% savings; no source given.
  • Typical ceiling on AI budgets. Until return is visible, he expects most businesses to stay at 0.1-0.4% of revenue. His estimate.
  • What it means for the build-out. Labs ordered capacity years ahead while customers learn to buy more efficiently. Jones: demand “can still grow astonishingly quickly and yet disappoint everyone who financed the equipment”. He says this is not evidence of overbuilding, and that he still thinks businesses are under-allocated to AI. The Wall Street Journal report he cites, that Jensen Huang told his finance team in August that committed wafers and memory were coming “like a quote avalanche”, is in the Nvidia CEO-transition article.
  • Where he expects value to move: up the stack, to products that use tokens to produce revenue, and to evals, routers and harnesses.

The pass-through problem for service sellers

  • Infosys, October earnings call (relayed by Jones): AI revenue was growing quickly while customers demanded that AI productivity gains be passed through in pricing; operating margin held around 21%. Not checked against the call.
  • His worked example: a service priced at 6,000 to deliver. AI cuts delivery to 2,000 of margin lasts until a competitor offers the same result for less.
  • For agencies and consultancies: ^[inferred] this is the same pressure Paul Roetzer describes for agencies selling AI-assisted content; see AI Transformation Case Studies.

What survives the next model

Jones’s rule for where to put AI investment:

Likely to lose valueLikely to gain value
Complicated harnesses built to work around a model’s weakness; he has seen several simplified this yearOrganised customer information
Plans that depend on a technical problem staying hardConnected systems
A reliable way to judge whether a result is correct

Work in the left column can still be worth doing if it pays back quickly.

Try It

Jones’s checklist for an AI spending decision in Q4 2026 or Q1 2027, as steps:

  1. Write the eval from real assignments, including the difficult edge cases.
  2. Set the pass standard before you test, as a baseline, not tuned to the model you prefer.
  3. Start with a relatively cheap model and move up only on a failure.
  4. Count the full cost: tokens, review time, correction time and any harness changes. A cheap model passes only if it wins on all of them.
  5. When a cheap model fails, read why. Either the failure shows what the expensive model is worth, or two or three cheap attempts still cost less.
  6. Note whether the race is close. If several models pass, you have a choice of suppliers and a case for a discount.
  7. Re-run the same eval on each new release instead of waiting to be told where it fits.
  8. Ask what survives the next model before funding a harness.

For Claude specifically, apply the test inside the lineup first: Which Claude Model for Which Job gives the starting tier, and Orchestrator + Cheap-Worker Routing gives the split.

Open Questions

  • One month or a trend? Ramp’s weekly dashboard (search snippet, not fetched) shows token spending up 10.5% in the week ending November 1 against the same week a month earlier, which sits oddly beside 1.65% for October. The two may measure different things (token spend against all AI spend).
  • Sample. Ramp’s customers skew toward startups and tech. A press write-up (search snippet) says the index covers more than 85,000 businesses.
  • Jones’s enterprise conversations are unnamed and cannot be checked.
  • The Antigravity Orchestrator figures are Google’s own, relayed by Jones; no Google page was fetched.
  • Jones mentions an evals kit on his Substack; it was not fetched.
  • Menlo’s survey in Enterprise AI Spend, November 2026 found spend per company up 4.7× year on year. That is a yearly survey figure and Ramp’s is a monthly transaction figure; they are not in conflict but should not be averaged.