In the current AI models space, there’s OpenAI and Anthropic, and then there’s everyone else.
Ramp’s latest AI Index update delivers a clear, weekly view of where businesses are actually putting their AI dollars. Drawing from its Token Spend Management product—which connects to provider accounts for granular usage and billing data across a sample of firms—the platform ranked models by token spend for the week of September 14–20, 2026. The tracked total came to $27.9 million (a sample figure; overall volumes are higher but directionally similar). Interestingly, the top 10 models were all either from OpenAI or Anthropic.
OpenAI’s newly launched GPT-6 Astra sits at the top.
The Top 10 by Token Spend
Here is the leaderboard, ranked by dollars spent on tokens (with week-over-week rank changes and share of the sample total):
- GPT-6 Astra (OpenAI) — $5.3M (18.9%) — NEW
- Claude Opus 5 (Anthropic) — $4.8M (17.1%) — down 1
- GPT-5.6 Sol (OpenAI) — $3.2M (11.3%) — down 1
- Claude Opus 4.8 (Anthropic) — $2.1M (7.6%) — down 1
- Claude Fable 5.1 (Anthropic) — $1.8M (6.6%) — NEW
- Claude Sonnet 5 (Anthropic) — $1.5M (5.2%)
- Claude Sonnet 4.6 (Anthropic) — $1.1M (3.8%) — down 2
- GPT-5.6 Terra (OpenAI) — $817.1K (2.9%) — down 1
- GPT-5.6 Luna (OpenAI) — $677.2K (2.4%) — up 4
- Claude Fable 5 (Anthropic) — $576.2K (2.1%) — down 6
OpenAI models account for the largest single share at the very top, driven by Astra’s rapid uptake. Anthropic, however, places six models in the top 10 and continues to hold a strong overall position in business adoption.
What the Ranking Shows
Astra’s debut at No. 1 is the clearest signal. The model has quickly captured nearly one-fifth of the sample’s token spend, reflecting both its premium pricing and strong demand for frontier capability. It arrives after months of internal demonstrations and public results on hard problems, including a batch of ten open mathematics challenges spanning sphere packing, group theory, and extremal combinatorics that OpenAI attributed to an internal version of the model.
Claude Opus 5 remains a close second, underscoring Anthropic’s continued strength with enterprises that prioritize its models for production workloads. Older but still heavily used models such as Claude Opus 4.8, GPT-5.6 Sol, and the Sonnet line show that businesses are not abandoning prior generations overnight. Cheaper or more specialized options like GPT-5.6 Luna and Terra also appear, illustrating the ongoing mix of high-end frontier spend and cost-conscious volume usage.
The ranking is based on observed API costs among businesses connected to Ramp’s Token Spend Management tool. It is not a complete market census, but it offers one of the more transparent, transaction-grounded views available of real customer spend rather than list prices or self-reported surveys.
Broader Context
Ramp’s wider AI Index continues to show paid AI adoption at 56.1% among the more than 70,000 U.S. businesses on its platform—well above Census Bureau estimates. Spend remains highly skewed: the median firm spends only around $12 per month, while the top 1% spend thousands per employee. Frontier models still command a large share of dollars even as effective token prices have fallen and some firms shift volume toward more economical options.
The data arrives as both labs continue releasing capable systems that push performance in coding, reasoning, and specialized domains such as mathematics. Models that solve long-standing open problems or deliver measurable productivity gains naturally attract enterprise budgets, and the current leaderboard reflects that reality in real time.
Ramp plans to expand the tokenomics section further, including more open-source coverage and team-level breakdowns. For now, the weekly snapshot makes one thing plain: customers are voting with their tokens, and GPT-6 Astra has taken the early lead.