Guanxi Technology Research · Week to 25 Oct 2026

AI token economics weekly

What businesses and developers are buying, what it costs, and what changed this week. The summary comes first, then six charts.

The picture in five points

1

A third of the tokens, 5% of the spend

On Vercel's AI Gateway, DeepSeek models account for about a third of the tokens processed but about 5% of estimated spend. Anthropic models account for 11% of tokens and 33% of spend. Among US businesses on Ramp, Anthropic accounts for 60% of token spend.

2

Open weights win on volume

Open-weight models account for 68% of tokens processed on Vercel's AI Gateway and 34% of its estimated spend (28% of estimated spend on OpenRouter). Part of this week's jump is Vercel reclassifying Beam 1.0 as open weight. Only 0.3% of US businesses on Ramp pay for DeepSeek.

3

Model choice is the cost lever

Summarizing 1,000 documents costs about $1.55 on the cheapest tracked model and about $155 on the most expensive. Of the 16 models we track, 15 had unchanged list prices this week.

4

Demand keeps growing

Tokens processed on OpenRouter rose 15% to 188T and estimated spend rose 10% to about $192M. Token spend among US businesses on Ramp rose 23% over the past month. Spend grows more slowly than tokens, which fits a shift to cheaper models.

5

Watch item: compute costs

GPU rental list prices have not moved across four captures since 10 Oct. One modeled index shows B200 up 5% and B300 up 7% this week. One index over three captures is not a trend.

What it means

These are GTR's interpretations of the data, not facts from it.

For enterprises buying

  • Cost per task is invisible in a pilot and decisive at scale. At list prices, 10,000 document summaries a month (about 10,000 words each) cost about $16 on the cheapest tracked model, $310 on Claude Sonnet 5.5 and $620 on Claude Opus 5.5. Price your real volumes before you pick a default model.
  • Set a default tier and make premium the exception. Across the platforms we track, much of the volume runs on cheaper models while a large share of the spend goes to premium ones. If most of your own spend is on premium models, check whether every task needs that tier. Start with high-volume routine work such as summaries and support replies.
  • Keep the model swappable. Vendor shares can move quickly. In one month Anthropic's share of token spend among US businesses on Ramp rose 6 points and OpenAI's fell 6. Put a routing layer between your apps and any one vendor, and avoid long fixed commitments at today's prices.
  • Do not count on falling prices. List prices were almost all flat this week. Savings so far come from moving work to cheaper models. Ask vendors for price protection and notice of changes.
  • Test open-weight models on your own workloads. Open-weight models account for about two thirds of tokens processed on Vercel and OpenRouter, and several cost far less per task than premium models (GLM 5.3 Flash is $2.20 per 1,000 summaries at list prices). Check license terms, where the model is hosted and how your data is handled. Self-hosting means renting GPUs ($2.89 to $7.53 an hour on the Silicon Data index), which usually makes sense only at steady, high volume.
  • Compare list prices with your invoices. Our spend figures are list prices. Your discounts, caching and batch rates change the real picture.

For investors following the market

  • Do not size the market by tokens. Volume leans toward cheap and open models, while a large share of spend goes to premium vendors. Anthropic models account for 11% of tokens on Vercel's AI Gateway but 33% of its estimated spend, and 60% of token spend among US businesses on Ramp.
  • Usage may be growing faster than revenue. On OpenRouter this week, tokens rose 15% and estimated spend rose 10%. If the mix keeps shifting to cheaper models, spend per token falls even with flat list prices.
  • Watch open-weight share of spend. On Vercel, open-weight models account for about two thirds of tokens and about a third of estimated spend. A rise in the spend share would suggest cheap models are starting to take more of the money. This week OpenRouter's open share of spend fell from 36% to 28%.
  • Compute costs are a watch item. Flat provider prices with firmer rental indices could squeeze providers who resell compute. One index and three captures is too little to act on.
  • Monitor weekly: open share of spend, vendor shares on Ramp, GPU index against list prices, and any list price changes.

Tokens vs. money, by lab

Share of tokens and spend on Vercel AI Gateway, with change vs last week.

Bar chart of Vercel AI Gateway token share and spend share by lab, 19 to 25 October 2026. DeepSeek 32% of tokens and 5% of spend. Anthropic 11% of tokens and 33% of spend. OpenAI 10% and 15%. Z.ai 20% and 20%.

Open weights: volume vs. money

Open-weight share of tokens and estimated spend, weekly, since 24 Aug.

Line chart of open-weight share of tokens (now 68%) and estimated spend (now 34% on Vercel) by week from 24 August to 19 October 2026.

1,000 summaries: what they cost

The same job at list prices, from Claude Haiku 5.5 at $1.55 to GPT-6 Astra at $155.

Bar chart of the list-price cost of 1,000 document summaries across nine models, from $1.55 to $155.

What changed this week

Week on week, 19–25 Oct vs 12–18 Oct.

List of week-on-week changes in model share, open-weight share, OpenRouter volume, estimated spend and list prices.

GPU rental prices

Provider list prices are flat. The Silicon Data index shows the newest chips firming.

Table of GPU rental prices per hour. H100 $2.89, H200 $3.47, B200 $6.60, B300 $7.53 on the Silicon Data index, with provider list price ranges.

What US businesses pay for

Ramp card data on token spend share and the share of businesses paying.

Ramp AI Index. Anthropic holds 60.1% of business token spend, OpenAI 36.4%. 43.8% of US businesses pay for Anthropic, 40.6% for OpenAI.

How to read this