Open-weight models capture 56% of AI Gateway token volume as Astra doubles Fable 5.1 spend

Open-weight models have reached a tipping point in production, now handling more volume than all closed-weight models combined. While they account for 56% of tokens, they only represent 14% of total spend. This indicates a strategy where teams use open-weight models for the bulk of their workloads and reserve frontier models for specific high-premium tasks.

The cost of inference is dropping rapidly, with the average token price falling 23.2% in August. For teams processing over ten million tokens, the median cost decreased by 7.6%. This trend allows operators to increase inference volume without increasing their existing budgets.

Market share for frontier models is volatile and driven by price-to-capability ratios. Fable 5 saw its spend share drop from 13.2% in July to 4.9% in August because Opus 5 offered similar utility at half the price. Similarly, GPT-6 Astra captured twice the spend share of Fable 5.1 despite launching at the same price point.

Lab loyalty is tied to model profiles rather than brand identity. Anthropic has maintained the top two spend positions since December by providing viable step-down options like Opus 5. In contrast, Gemini 3 Flash lost 95% of its token share since May, with most of that volume moving to competing labs.