Claude Fable 5.2 and Next.js 16.4.6: what to check first
Dev news, 3 to 10 Nov 2026: Claude Fable 5.2 costs less but breaks code that edits history, Next.js 16.4.6 compresses route handlers, Vercel builds get 64 GB.
The week's main change is Claude Fable 5.2, released on 9 November 2026: it costs a fifth less than Claude Fable 5.1 per token, and it breaks code that edits a conversation's history, still uses the computer_20251124 tool or pairs the advisor tool with an older model. Before you switch, check those three things and whether your team runs under zero data retention, which Fable 5.2 does not allow. Next.js 16.4.6, out the same day, backports three fixes and its notes name no advisory, but its route handler compression fix also strips the Content-Encoding and Content-Length headers from the responses route handlers return, citing an advisory from the Next.js 16.4.5 security release: if a route handler passes on a fetch() response, test it after the upgrade. On Vercel, the Workflow SDK writes over WebSockets from workflow 5.2.0, builds on Basic and Standard machines get 64 GB of disk, Sandbox snapshots are up to 10 times faster and the Container Registry takes 2 GB layers.
| Item | Version or date | What to do |
|---|---|---|
| Claude Fable 5.2 | claude-fable-5-2, 9 Nov | Check history edits, computer use and the advisor first |
| Claude Fable 5.2 on AI Gateway | anthropic/, 9 Nov | Check your team's ZDR setting; add a fallback list |
| Next.js 16.4.6 | 9 Nov, three backported fixes | Test route handlers that return fetch() or stream |
| Workflow SDK over WebSockets | workflow 5.2.0, 9 Nov | Upgrade; on 4.x read the v5 guide first |
| Build disk on Basic and Standard | 32 GB to 64 GB, 9 Nov | Nothing; retry builds that ran out of disk |
| Sandbox snapshots | 4.4 s to under 0.4 s, 9 Nov | Nothing; new snapshots get the speed |
| Container Registry layers | 500 MB to 2 GB, 9 Nov | Stop splitting large layers; 15 GB per image still applies |
| AI Gateway Production Index | October data, 9 Nov | Price a decision model for your yes-or-no calls |
| GLM 5.3 Flash in Token Pass | 9 Nov | Pro teams only; $30 a month with $80 of credit |
Claude
Claude Fable 5.2: cheaper per token, with four breaking changes
On 9 November 2026 Anthropic released Claude Fable 5.2 (claude-fable-5-2), per the Claude release notes (opens in a new tab). It costs $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Claude Fable 5.1, has a 1M token context window and 128k output tokens, and thinks adaptively on every request. Fast mode (opens in a new tab) for it costs $16 and $80 and needs only speed: "fast", no beta header. Like Fable 5.1 it requires 30-day data retention and isn't available under zero data retention unless Anthropic expressly authorises it.
What's new in Claude Fable 5.2 (opens in a new tab) lists four changes that can break code written for Fable 5.1. The first is the prefix check: a thinking block the model produced is valid only while everything before it (the system prompt, the tools and every earlier message) is unchanged. Fable 5.1 returns that error by default only on accounts created on or after 31 August 2026; Fable 5.2 returns a 400 error saying "The block is bound to a different conversation" on every account. Second, only Fable 5.2 and Mythos 5.2 read Fable 5.2's thinking blocks, so a router that moves a conversation to Opus 5.5 or Fable 5.1 loses that reasoning (the request still succeeds). Third, the computer_20251124 tool returns a 400 everywhere except Amazon Bedrock; declare the computer_toolset_20260801 toolset instead. Fourth, with the advisor tool (beta), a Fable 5.2 executor accepts only a Fable 5.2 or Mythos 5.2 advisor.
Our take: the prefix check is the one that bites, because code that ran clean on an older account can fail only after the switch. Search your code for anything that rewrites system per request (a date, a mode flag), trims old tool results or reorders messages; each is an edit. The drop_block setting under the thinking-binding-controls-2026-08-01 beta header stops the errors, but Anthropic warns it can raise output tokens, so use it to ship while you make the history append-only. Anthropic's page also says to start most workloads on Opus and move to Fable 5.2 only when your evals at higher effort fall short: the price cut makes Fable cheaper, not the default.
Claude Fable 5.2 on Vercel AI Gateway
Vercel added Claude Fable 5.2 to AI Gateway (opens in a new tab) on 9 November 2026 as anthropic/, callable through the AI SDK, Chat Completions, Responses and Anthropic Messages APIs. The model requires data retention: prompts and completions are kept for 30 days and not used for training, and a team with Zero Data Retention enabled gets an error instead of a response. Cybersecurity and biology classifiers are on, and Vercel recommends a models fallback list under providerOptions.gateway so that a request the cybersecurity classifier declines moves to another model; its example chain is Claude Opus 5.5, then Sonnet 5.5.
Our take: note the ID, with a dot on the Gateway and hyphens on the Claude API, when you move code between the two. Set up the fallback list, but know what it costs: a conversation that falls back from Fable 5.2 to Opus 5.5 continues without Fable's thinking blocks, so the next turn starts its reasoning over. If your team enabled ZDR on the Gateway for a client, Fable 5.2 is simply off the table for that team.
Next.js
Next.js 16.4.6: compressed route handlers, stripped headers
Next.js 16.4.6 (opens in a new tab) was published on 9 November 2026 as a backport of bug fixes to the 16.4 line. It compresses route handler responses on self-hosted servers, pins Tailwind to 4.3.3 in create-next-app, and brings Turbopack garbage collection fixes for next dev. The release notes name no security advisory, but the compression pull request (opens in a new tab) does. It explains the fix: on a self-hosted server, App Router route handler responses, prerendered ones included, were never compressed, because a header value stored as an array made the compressible check skip them. The same pull request strips the Content-Encoding and Content-Length headers from the Response a route handler returns, citing GHSA-pc58-hwjq-p385, one of the seven advisories fixed in 16.4.5, in which a forwarded fetch response kept compressed framing headers on a decompressed body. Route handlers should return responses uncompressed and leave compression to the server that sends them, next start or a host like Vercel, and the pull request says a route handler should not return a fetch() response without processing its body.
Our take: if a route handler proxies an upstream API by returning its fetch() response as is, check after the upgrade which headers your clients get, or better, read the body and build a new Response, as the pull request advises; for a real proxy it points to a dedicated proxy library. If you run next start and no proxy in front of it compresses, JSON from your route handlers now leaves gzipped, a free saving on large payloads. If a route handler streams, such as server-sent events for a chat, test that the chunks still arrive one by one. And if you skipped 16.4.5, read its advisories first.
Vercel
The Workflow SDK writes over WebSockets
From workflow 5.2.0 (opens in a new tab) (9 November 2026), the Workflow SDK sends event writes and stream writes to Vercel Workflows over a WebSocket instead of one HTTP request per write. It needs no code change beyond the upgrade; new runs take the WebSocket path after you deploy, and runs in flight finish on the deployment they started on. In Vercel's benchmark, replaying an AI agent's token stream, 5.2.0 delivered the full response 4.2 times faster than 4.8.12.
Our take: every step's start and finish is an event write, so a workflow with many short steps gains the most. On 5.x, upgrade; on 4.x, this sits behind a major version, so read the v5 upgrade guide before you chase the speed-up.
Build disk on Vercel grows to 64 GB
Builds on Basic and Standard machines now get 64 GB of disk, up from 32 GB (opens in a new tab), on every plan, at no extra charge and with no configuration, per the changelog of 9 November 2026. Enhanced and Turbo machines already had 64 GB.
Our take: if a monorepo build failed with "no space left on device" and you moved to Enhanced machines only for the disk, try Standard again; you may not need to pay for the CPU you weren't using.
Vercel Sandbox snapshots are up to 10 times faster
On 9 November 2026 Vercel made Sandbox snapshot creation up to 10 times faster (opens in a new tab): the median snapshot takes under 0.4 seconds, down from 4.4, and the median restore 250 ms, down from 500. New snapshots get the new format automatically; existing ones keep the old one.
Our take: persistent sandboxes, the default, snapshot whenever a session stops and resume from the latest snapshot, so an agent that pauses and resumes many times a session gains the most. The speed shows from each sandbox's next snapshot, not on the ones you already have.
Vercel Container Registry takes 2 GB layers
The Vercel Container Registry (opens in a new tab) accepts compressed image layers of up to 2 GB since 9 November 2026, up from 500 MB; an image is still capped at 15 GB. The vercel vcr commands need Vercel CLI 58.3.0 or later.
Our take: a layer with model weights or a GPU runtime no longer has to be split into several layers to fit under 500 MB. The 15 GB limit per image is now the one to watch.
AI Gateway
Decision models take a quarter of Gateway requests
Vercel's AI Gateway Production Index for November 2026 (opens in a new tab), published on 9 November and built on anonymised Gateway traffic through October, finds that decision models, models that "classify, score, or pick between choices instead of generating text", took 26.1% of requests in October but 0.53% of spend. Close to a third of paid teams used one. Anthropic was the first-choice lab of 42.7% of new teams in October (34.4% in September), OpenAI of 24.0%.
Our take: these are Vercel's numbers about its own traffic, not a market share, but the gap between requests and spend is the useful part. If a language model answers yes or no in your app (spam, routing, whether a link fits), price a decision model for that call. We do this ourselves: Jev, the TypeSafe model the index names as the one that got almost all of those requests, is the judge that checks the links on this blog.
GLM 5.3 Flash joins Token Pass
Vercel added GLM 5.3 Flash to Token Pass (opens in a new tab) on 9 November 2026: Z.ai's multimodal coding model with a 1M token context window, next to Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash. Token Pass is a subscription through AI Gateway for Pro teams, $30 a month with $80 of monthly credit, with Zero Data Retention on by default.
Our take: it is the opposite trade to Fable 5.2: a fixed monthly price and zero retention, with Z.ai's coding model in place of Fable. For a coding agent on a client's code that must not be retained, that default settles the question before the benchmarks do.
What changed in Claude Fable 5.2?
Claude Fable 5.2 (claude-fable-5-2), released on 9 November 2026, costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Claude Fable 5.1, with a 1M token context window and 128k output tokens. Four changes can break code: the prefix check returns a 400 error on every account, only Fable 5.2 and Mythos 5.2 read its thinking blocks, the computer_20251124 tool returns a 400 outside Amazon Bedrock, and a Fable 5.2 executor in the advisor tool needs a Fable 5.2 or Mythos 5.2 advisor.
Claude Fable 5.2 vs Opus 5.5: which should I use?
Anthropic's own advice is to start most workloads on the latest Claude Opus model, Opus 5.5, and move to Claude Fable 5.2 only when your evals on Opus at higher effort still fall short. Fable 5.2 is built for demanding reasoning and long agentic work, requires 30-day data retention, and a conversation that moves from Fable 5.2 to Opus 5.5 loses Fable's thinking blocks.
Why does Claude Fable 5.2 return 'The block is bound to a different conversation'?
That is the prefix check. A thinking block stays valid only while everything before it, the system prompt, the tools and earlier messages, is unchanged. If your code edits any of them and sends an older Fable 5.2 thinking block back, the request returns a 400 error on every account. Keep the history append-only, or as a stopgap send the thinking-binding-controls-2026-08-01 beta header with thinking.block_binding.prefix_mismatch_behavior set to drop_block.
Can I use Claude Fable 5.2 with zero data retention?
No, unless Anthropic expressly authorises it. Fable 5.2 requires 30-day data retention, like Fable 5.1. On Vercel AI Gateway, a team with Zero Data Retention enabled gets an error instead of a response; prompts and completions are kept for 30 days and not used for training.
Is Next.js 16.4.6 a security release?
Not by its release notes, which name no advisory. Next.js 16.4.6, published on 9 November 2026, backports three bug fixes: compression for route handler responses on self-hosted servers, Tailwind pinned to 4.3.3 in create-next-app, and Turbopack garbage collection fixes in development. But the compression pull request also strips the Content-Encoding and Content-Length headers from the Response a route handler returns, citing GHSA-pc58-hwjq-p385, an advisory fixed in Next.js 16.4.5 on 3 November, and says a route handler should not return a fetch() response without processing its body.