Simon Willison reports Qwen3.6-35B outperforms Claude Opus 4.7 on SVG art; Qwen 3.8 27B described as nearly competitive with frontier models
Concluded·Day 6·in the edition since 22 Sep 2026·6 sources·updated 27 Sep 2026
Chinese open models' lead over US labs
The gist
Open-weight models at 78% of AI tokens; closed models hold most revenue
Token volume and revenue are moving in opposite directions, with open-weight models capturing usage while closed frontier models retain pricing power. Goldman Sachs data shows frontier token demand growing slower than total demand, which affects how AI infrastructure investment can be sustained as average token prices fall.
The full picture
Vercel AI Gateway data shows open-weight models at 78.4% of token volume as of September 2026, up from roughly 40% in June. Revenue runs in the opposite direction: DeepSeek V4.1 Flash generates 59.3% of tokens but only 5.1% of spending, while Anthropic's five listed models collectively capture 36.9% of total spending despite a small token share. SignalPlus, reporting as of late August, placed open-weight token share at 62% and Anthropic's spending share at 61 to 65 percent. Open-weight models attract volume by running at roughly one-tenth the price of closed models and allowing developers to self-host, fine-tune, and avoid API lock-in.
Goldman Sachs research, cited by Rohan Paul, adds an infrastructure dimension: total token demand grew approximately 18x from December 2025 to September 2026, but frontier-model token demand grew only 8 to 9x. Efficiency techniques such as quantization, distillation, speculative decoding, and model routing reduce compute per token, so GPU demand can grow far slower than raw token volume. Goldman Sachs said the relevant infrastructure metric is tokens consumed multiplied by compute required per token, not token volume alone.
At the company level, MilkRoadAI reported OpenAI at approximately $40 billion in annual recurring revenue and Anthropic at approximately $65 billion, while all major Chinese AI developers combined generate approximately $10.7 billion. DeepSeek generates an estimated $500 million in annual recurring revenue; ByteDance leads Chinese AI revenue at approximately $4 billion, Alibaba at $2.4 billion.
OpenAI cut GPT-5.6 Sol input pricing from $5 to $4 per million tokens and output from $30 to $20 per million tokens as a three-month promotional offer. OpenAI had already cut prices for smaller models, with GPT-5.6 Terra down 20% and Luna down 80%. Anthropic's Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens; Claude Opus 5 at $5 and $25.
How it developed
Goldman Sachs research published September 26 found total AI token demand grew approximately 18x from December 2025 to September 2026 while frontier-model demand grew only 8 to 9x, with open-weight and cheaper models absorbing the marginal growth.
The research added that the relevant infrastructure metric is tokens multiplied by compute per token, not volume alone, as efficiency techniques including quantization and distillation reduce compute per token. OpenAI cut GPT-5.6 Sol input pricing from $5 to $4 per million tokens and output from $30 to $20 as a three-month promotional offer.
Treasury Secretary Scott Bessent calls for more US open-source AI models, warning against regulatory capture by large labs and framing open-source development as a counter to China.
NVIDIA AI Day Singapore showcases Southeast Asian organizations customizing Nemotron open models
Epoch AI's JS Denain estimates Chinese labs trail US frontier labs by six to eight months on public releases
First in the edition · 22 September 2026
A briefing for U.S.
Congressional members, published September 22, finds Chinese open-weight models score 42-45 on the Artificial Analysis Intelligence Index against 23-26 for American open models, hold a 1.6 billion Hugging Face download lead, and appear in over 40% of AI/ML arXiv papers against roughly 30% for U.S. models; distillation of American frontier models accounts for only an estimated 1-2 months of the gap. Vercel AI Gateway data shows open-weight models at 78.4% of token volume, with DeepSeek V4.1 Flash at 59.3% of tokens but only 5.1% of spending.
First citedMilk Road AI Twitter
Alibaba releases Qwen-Image-2.1, a 7B open-weight image generation and editing model with ComfyUI and vLLM-Omni support.
MindStudio publishes analysis showing open-weight models surpassed closed models in Vercel token share, with DeepSeek passing Anthropic at 25.2% vs 24.5%
OpenAI cuts GPT-5.6 Sol input pricing from $5 to $4 and output from $30 to $20 per million tokens for three months
SignalPlus reports open-weight models at 62% of Vercel AI Gateway token traffic, with Anthropic commanding 61-65% of spending
Sources
Milk Road AI Twitter
Rohan Paul Twitter
1 more source
Related
The daily email
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free