September 2026 Summaries

AI Gateway’s September 2026 Production Index, based on anonymized traffic through August, reports that open-weight models accounted for a majority of token volume for the first time, rising from 7% in December 2025 to 56% in August while representing 14% of spending. This shift, alongside broader model-price reductions, contributed to a 23.2% decline in average token prices in August, with the median high-volume team paying 7.6% less per token. Anthropic retained 64% of gateway spending despite Fable 5 losing substantial share after the less expensive Opus 5 launched, as many customers shifted within Anthropic’s portfolio rather than changing providers. Google’s Gemini 3 Flash saw its token share decline sharply, with much of its displaced volume moving to competing providers’ cheaper or higher-capability models. In early September, OpenAI’s GPT-6 Astra quickly captured a major portion of OpenAI spending and exceeded Anthropic’s similarly priced Fable 5.1 in early adoption and gateway spend. The report also notes changes in multimedia generation, including Google’s Nano Banana leading image spending, Seedance maintaining leadership in video generation and spending, and declining share for xAI’s Grok Imagine.

Vercel redesigned how its CDN retrieves routing metadata for deployments, replacing separate per-path cache objects with bounded, indexed metadata shards that warm the cache for many routes at once while allowing individual records to be found through binary search without parsing entire files. The change addresses frequent cache misses in large, regularly deployed projects, where new deployments previously generated fresh metadata keys for potentially hundreds of thousands of paths. Production testing identified roughly 200 KB as the best shard size, balancing transfer costs, process-level cache behavior, and regional cache hit rates; the approach reduced P99 metadata lookup latency from 215.8 ms to 19.1 ms and average latency by 79%. Vercel validated the migration through offline comparisons and feature-flagged shadow mode, which also exposed a rare legacy path-encoding bug involving split emoji. Removing now-redundant per-path metadata build work saved about 16.6 seconds per deployment, making deployment steps about 10% faster overall and an estimated 25% faster for metadata-heavy projects, while preserving the existing Build Output API contract.

Vercel has introduced Flat Rate CDN for Pro teams to replace unpredictable usage-based CDN charges with fixed monthly pricing, spike protection, and capacity tiers based on typical traffic levels. New Pro teams receive the feature by default, while existing teams can opt in; coverage applies across all projects within a team and includes CDN requests, fast and Blob data transfer, and CDN-generated observability events. Plans range from an included tier offering 1 million requests and 1 TB of transfer to paid tiers supporting 10 million, 50 million, or 150 million requests with 50 TB of transfer, priced at $20, $100, and $300 per month respectively. Temporary traffic surges do not cause overages, throttling, degraded performance, or service interruption, as Vercel excludes spikes when assessing capacity and may adjust tiers for sustained growth in later billing cycles. The company says the service continues to use its premium network infrastructure, provides usage monitoring and capacity alerts, and allows teams to opt out and return to pay-as-you-go CDN billing.

Vercel describes Fluid as a unified compute system designed to rapidly assemble tailored, isolated environments for builds, functions, sandboxes, and agent workloads, reducing the need for developers to manage separate infrastructure primitives. Built on Hive for VM provisioning, Fluid Images for fast-resuming custom environments, and Vercel Drives for portable persistent storage, the platform aims to match compute resources to workloads ranging from CPU-intensive builds to I/O-focused functions and flexible agent sandboxes. Vercel says Fluid already powers more than 15 million daily builds, 25 million weekly sandboxes, and a trillion monthly requests, while allowing workloads to share instances and charging Active CPU pricing only when code is actively using CPU. The company presents the system as particularly suited to agents, which require immediate provisioning, strong isolation for untrusted code, customizable tooling, and state that persists across frequently replaced machines.