Vercel AI Updates: November 10, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Claude Fable 5.2 Is on AI Gateway With Fallbacks for Declined Cyber Requests and No Zero Data Retention Support
Vercel added Anthropic’s Claude Fable 5.2 (anthropic/claude-fable-5.2) to AI Gateway on November 9. Vercel describes it as built for open-ended technical work that can run long, ambiguous tasks unattended, usable as an orchestrator that delegates to smaller models, with effort levels from low to max and high as the API default. Anthropic does not support Zero Data Retention for the model, so prompts and completions are retained for 30 days without being used for training, and AI Gateway returns an error instead of sending the request when ZDR is enabled for a team. Cybersecurity and biology safety classifiers are on, and if the cybersecurity classifier declines a request, the gateway can retry with the models listed in providerOptions.gateway.models, in order. Vercel states pricing passes through at provider rates with no markup and gives no per-token figures or context size; the model works through the AI SDK, Chat Completions, Responses and Anthropic Messages APIs. Source
2. Meta’s Muse Ember Joins AI Gateway as a Lightweight 1M-Token Reasoning Model for High-Volume Agent Steps
Vercel added Meta’s Muse Ember (meta/muse-ember) to AI Gateway on November 9, described as a lightweight reasoning model for the high-volume parts of an agent run such as coding sub-tasks, tool calls and document analysis. It accepts text, image and video input, has a 1M-token context window, supports tool calling, and offers adjustable reasoning effort from minimal to max. It can be reached through the AI SDK, OpenAI-compatible Chat Completions and Responses APIs, the Anthropic Messages API, and coding agents connected to the gateway, and can be tried in the model playground. The announcement lists no Muse Ember-specific rates, only that the gateway passes provider pricing through without markup. Source
3. GLM 5.3 Flash Is Added to Token Pass, the $30 per Month Subscription With $80 in Credit
Vercel added Z.ai’s GLM 5.3 Flash, a native multimodal coding model with a 1M-token context window, to Token Pass on November 9. Token Pass is a Pro-team subscription that runs through AI Gateway at $30 per month with $80 in monthly credit and Zero Data Retention on by default, and it can be used with coding agents such as fx, Pi, Cline and Command Code or with the AI SDK. GLM 5.3 Flash joins Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash in the plan, and Vercel says more models are coming. Teams add it from the Token Pass page in the AI Gateway dashboard. Source
4. Vercel Sandbox Snapshot Creation Drops From 4.4 Seconds to Under 0.4 Seconds at the Median
Vercel Sandbox now creates snapshots up to 10x faster, with median creation time under 0.4 seconds, down from 4.4 seconds, and restore time at p50 down from 500 ms to 250 ms. The gain comes from writing and restoring snapshots by reusing frames in the Vercel Hive Snapshot (VHS) format, which indexes data so more reads and writes run in parallel and lets restores load more of the base image alongside user data. Persistent sandboxes, the default, snapshot automatically when a session stops and resume from the latest snapshot, so the change shortens stop and resume cycles for agent workloads. It applies to all new snapshots with no configuration, while existing snapshots keep the previous format. Source
5. Workflow SDK 5.2.0 Writes Events and Streams Over WebSockets, Delivering an Agent Token Stream 4.2x Faster
Vercel changed the Workflow SDK to send event writes and stream writes to Vercel Workflows over a WebSocket by default, instead of one HTTP request per write (or per write group for streams). It requires workflow@5.2.0 or later with no code change beyond upgrading, and teams on 4.x need to follow the v5 upgrade guide because v5 is a major release. In a benchmark replaying an AI agent’s token stream, v5.2.0 delivered the full response 4.2x faster than v4.8.12, with no chunk delayed by more than one second. Every step start and completion is an event write, so workflows with many steps make fewer HTTP requests; new runs use the WebSocket path after deployment while in-flight runs finish on their original deployment. Source
6. Vercel’s November AI Gateway Index Reports Decision Models at 26.1% of Requests and a Six-Week Model Half-Life
Vercel published its AI Gateway production index on November 9, based on anonymized traffic through October 2026 across the gateway, which it says exceeds 100 trillion tokens per month. Decision models rose from 0.0% of requests in August to 12.8% in September and 26.1% in October while making up 0.53% of spend, and about a third of paid teams used one in October. Among new paid teams in October, 42.7% first chose Anthropic, 24.0% OpenAI, 16.5% Google and 6.6% DeepSeek, although Anthropic’s share of spend fell from 69.2% in December 2025 to 32.2% and OpenAI’s rose from 8.6% to 14.5%. Models 60 days old or newer drew 52.7% of language-model requests, about half of teams with at least 10 million monthly tokens changed their most-used model in October, and the report estimates a model half-life of about six weeks. Voice model use among paid teams rose from 4.1% in July to 10.7% in October. Source