Open-Weight Models Now Account for 71.5% of Tokens: The Enterprise AI Cost Landscape Is Shifting

A recent productivity index report from AI Gateway has drawn widespread industry attention. According to the data, open-weight models now account for more than 70% of the token traffic processed by AI Gateway, while their share of total spending has climbed to nearly a quarter. This shift signals a profound structural adjustment in enterprise AI deployment strategies.

The Dual Penetration of Open-Weight Models: A Leap in Both Volume and Value

According to the report, open-weight models currently process over 71.5% of gateway tokens and account for 23% of spending. Since June, their spending share has more than quadrupled. One detail worth noting: growth in open models is coming from both ends of the price spectrum — the cheapest models are driving token volume growth, while expensive models are pulling up spending.

High-priced open models like Kimi K3 cost at least four times the open-source average per token, yet contribute roughly two-thirds of open-model spending growth. In sharp contrast, DeepSeek^Ad-v4.1-Flash alone handled nearly half of September's token volume but accounted for only 2.6% of spending, with a per-token price just one-fifth of the open-source average.

This divergence shows that enterprise investment in open models isn't simply cost substitution — it's tiered configuration based on task requirements. Lightweight tasks pursue maximum cost-efficiency, while complex reasoning tasks are willing to pay for high-performance open models.

What Are "Open-Weight Models"?

Open-weight models are AI models whose parameters (weights) are publicly released, allowing anyone to download, deploy, and (in most cases) fine-tune them. They are not entirely equivalent to "open-source software": many open-weight models do not disclose training data, training code, or full training details — only final weights and inference code licenses. Typical examples include Meta's Llama series, the DeepSeek series, Moonshot's Kimi, and Z.ai's GLM.

For enterprises, the core value of open-weight models lies in self-hostability and customizability: companies can deploy models on their own infrastructure or third-party clouds, fine-tune them for specific tasks, and avoid lock-in risk from a single API provider. Meanwhile, companies like Mistral and Cohere also adopt a hybrid "open weights + commercial services" model. It's worth noting that the "openness" of open weights varies greatly — some licenses impose additional restrictions on commercial use, redistribution, or derivative models, so enterprises need to carefully review licensing terms when selecting models.

The Market Pie Is Expanding Rapidly: Spending Up 13x, Unit Prices Halved

The report reveals an even grander trend: the entire AI market is expanding at an astonishing rate while unit costs are falling fast. September gateway spending was approximately 13 times that of December, up 1,180%, while average per-token prices dropped 47%.

Median paying customers active in both December and September more than doubled their spending. This means enterprises aren't just using more AI — they're willing to pay more for more complex AI applications. That said, model and lab switching rates rose for the fifth consecutive month, with nearly half of teams changing their most-used model and nearly a third changing their most-used lab, setting an index record.

This high-frequency switching behavior indicates that models are becoming replaceable components within larger enterprise investments. Companies can flexibly swap underlying models while continuously improving AI products and workflows, without forming excessive dependence on any single lab.

A New Path for Chinese Models Going Global: US Providers Are the Main Beneficiaries

One of the report's most insightful findings is the emerging symbiotic relationship between Chinese open models and US cloud service providers. In September, more than two-thirds of gateway spending on China-origin models flowed to US-based providers — a dramatic increase from less than one-tenth in December.

Specifically, DeepSeek's models processed 58% of all gateway tokens but accounted for only 6.6% of spending. More than four-fifths of those tokens were served by DeepSeek itself, but two out of every five dollars in spending went through US providers. Moonshot AI (developer of Kimi) captured 3.6% of tokens and 13.2% of spending, with 90% of its tokens served by US providers. Z.ai (creator of GLM) accounted for 7.2% of tokens and 4.5% of spending, with 70% served by US providers.

For every $2.50 spent on Chinese models through US providers, only $1 is spent through Chinese domestic providers. This means the same Chinese model can compete with American models while generating revenue for American providers — a new value chain is taking shape.

What Is AI Gateway, and Why Is It a Window into the Market?

AI Gateway is a middleware service that sits between applications and multiple model APIs. Typical functions include: unified API formats, routing and load balancing, failover, cost tracking, rate limiting, and observability. Once enterprises connect through a gateway, they can switch underlying models or providers without changing application code — this is also the foundation that enables precise tracking of "model switching rates" in the report.

Representative products include Cloudflare AI Gateway, Portkey, LiteLLM, and OpenRouter. Because gateways can simultaneously observe token volume, spending amounts, model selection and other dimensions, they reflect the real multi-model usage landscape of enterprises better than any single lab's public data. Therefore, while the productivity index from gateways is not a full-market census, it is a high-quality sample for observing enterprise AI procurement behavior. Conclusions in this report — such as open-source share and provider flows — should be understood within this sample framework.

September New Model Scorecard: Jev's Rapid Adoption and Opus 5.5's Revenue-Generating Power

The report provides a multi-dimensional assessment of models released in September, revealing significant differences in adoption rates, spending shares, and retention rates.

TypeSafe AI's Jev became the fastest-adopted model in AI Gateway history, reaching 35% of paying teams within the first two weeks — nearly double that of second-place Sonnet 5.5. Jev is also a decision model: compared to typical language model requests, it's faster, cheaper, and uses very few tokens. As a result, despite its rapid adoption and broad reach, it accounted for only 1% of the month's tokens and 0.15% of spending.

Anthropic's Opus 5.5 captured over one-tenth of gateway spending — more than double any other September release. Nearly 90% of its spending flowed to teams already using Opus. However, Anthropic lost nearly 20 percentage points of gateway spending share in September, with the Opus series contributing 17 points. Anthropic's overall share fell below half of total gateway spending for the first time, dropping from 63.7% to 43.9%.

OpenAI's GPT-6 Luna reached nearly three-twentieths of paying teams, but had the highest retention rate of the month — nearly two-thirds of teams that tried it used it again the following week. In contrast, GPT-6 Astra reached fewer teams but captured more budget: Luna reached nearly three times as many teams as Astra, while Astra's spending share was more than five times Luna's.

Tokens and the "Token Volume ≠ Spending" Relationship

A token is the smallest billing unit for text processed by a model. Roughly, 1 English word equals about 1.3 tokens; Chinese is typically segmented by character or byte. API billing generally prices "input tokens + output tokens" separately, with output usually being more expensive.

To understand apparent contradictions in the report — such as "Jev accounts for 1% of tokens but only 0.15% of spending" or "DeepSeek accounts for 58% of tokens but only 6.6% of spending" — the key lies in task type and unit price differences: lightweight tasks like classification, routing, and short Q&A consume few tokens at low unit prices; long-context reasoning, code generation, and agent orchestration involve large token volumes at high unit prices. Therefore, token share measures "processing volume," while spending share measures "economic value." The divergence between the two is not contradictory — it actually reveals enterprises' strategy of tiering model配置 by task. This is also one of the report's most important analytical perspectives.

Overall Assessment and Trend Judgment

Several key trends can be distilled from this report. Open-weight models are moving from edge experiments to core enterprise workloads, and their steadily climbing token share indicates that technical maturity has crossed a critical threshold. Total AI market spending is growing rapidly, but unit prices are falling fast — typical characteristics of scale and technology diffusion. The continued rise in model switching rates signals that enterprise AI strategies are evolving from "vendor lock-in" to "multi-model orchestration."

For developers, this means flexibility in model selection and cost optimization capabilities will become increasingly important. For AI labs, this means simultaneously addressing competitive pressure from open models and declining customer loyalty. And for Chinese AI companies, the path to commercialization through US providers has become clear — this may emerge as an important model for global deployment.

One-line Summary

AI Gateway report shows open-weight models capture 71.5% of tokens and 23% of spending, as enterprise AI enters the era of multi-model orchestration.

Paragraph Summary

A productivity index report from AI Gateway shows open-weight models now account for 71.5% of gateway tokens and 23% of spending, with spending share quadrupling since June — growth coming from both the cheapest and most expensive ends. September gateway spending was 13 times December's, yet average per-token prices fell 47%, and model and lab switching rates rose for the fifth consecutive month. The report also reveals a symbiotic relationship between Chinese open models and US cloud providers — over two-thirds of Chinese model spending comes through US providers. Among September releases, TypeSafe AI's Jev set the fastest adoption record, while Anthropic's Opus 5.5 captured over 11% of spending but Anthropic's overall share fell below half for the first time. This marks a shift in enterprise AI strategy from vendor lock-in to multi-model orchestration.