Vercel this week: eve 1.0, Blob TTLs, Standard builds
Dev news, 11 to 18 Nov 2026: eve 1.0 is stable with breaking changes, Vercel Blob deletes files by TTL, new teams pay per minute for Standard builds on Vercel.

The week's main change on Vercel is eve 1.0, the open-source agent framework, generally available since 17 November 2026 with breaking changes from the preview: read the upgrade guide before you deploy it. The same day, Vercel Blob got TTLs that delete files by themselves (@vercel/ 2.9.0), new Pro and Enterprise teams started paying $0.014 a minute for Standard builds, and a new preset builds previews only for pull requests. AI Gateway added Private Inference for your own model endpoints and six catalog models, and on 16 November Anthropic extended its text watermark to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, after Claude Mythos 5, Opus 4.7 and Sonnet 4.6 got the text watermark.
| Item | Version or date | What to do |
|---|---|---|
| eve 1.0 | 17 Nov; eve 1.0.2 on npm, Node.js 24+ | Read the upgrade guide; upgrade servers and callers first |
| Standard builds for new teams | 17 Nov; $0.014 per minute on Pro | Budget builds for any team you create from now on |
| Pull-request-only previews | 17 Nov | Turn it on where branch previews nobody opens cost minutes |
| Vercel Blob TTLs | 17 Nov; @vercel/ 2.9.0, 1–730 days | Replace cleanup cron jobs with ttlDays |
| Private Inference on AI Gateway | 16 Nov; Vercel CLI 65.16.0 | Route self-hosted models through the Gateway; mind fallbacks |
| Decision-1.1 | 17 Nov; microsoft/ | Swap the ID, then re-tune your thresholds |
| GPT-6.2 Astra and Sol | 17 Nov | Rerun evals; check prices on the model page |
| Hy4 Flash at 50% off | 16 Nov, until 29 Nov | Run cheap evals now; don't budget on the discount |
| MAI-Image-2.7 and Flash | 17 Nov; AI SDK 6+ | Try transparent backgrounds for product images |
| Claude text watermark | 16 Nov; four more models | Nothing; omit sampling settings if you rely on it |
Vercel
eve 1.0 is generally available
Vercel announced eve 1.0 (opens in a new tab) on 17 November 2026, after a public preview that began in June. eve is an open-source framework for building agents: it is the harness, the loop that calls models, runs tools and delegates work to subagents, while you pick the model, the tools, the instructions and which actions need approval. 1.0 adds native integrations for Next.js, Nuxt, SvelteKit, TanStack Start and Vite, multi-tenant memory and credentials, and OpenTelemetry tracing. On npm (opens in a new tab) the latest version was 1.0.2 when we checked on 18 November, and it requires Node.js 24 or newer.
Our take: the upgrade guide (opens in a new tab) is the part to read. Extensions built with 0.83 or earlier are refused until you rebuild them, the session event stream moves from format v26 to v27 with most event names changed, and remote agent calls now use protocol 3. Callers and remote agents on different protocols can still exchange results, but a protocol-3 remote agent can't relay approvals, questions or sign-ins to a protocol-2 caller, so upgrade callers before the remote agents they call; calls to or from eve 0.66 to 0.68 fail at the start with REMOTE_AGENT_PROTOCOL_MISMATCH. Deploy the upgraded server before the upgraded client, and finish or end sessions still on 0.81.0 or earlier first, because they never move to the new deployment. From 1.0 on, eve follows semantic versioning, so stable APIs break only in a major release; that is what makes this the version to build a product on.
Standard builds move to usage-based billing for new teams
From 17 November 2026, new Pro and Enterprise teams no longer include a Standard build slot (opens in a new tab): their Standard builds are billed by the minute like the other build machines, at $0.014 per build minute on Pro, with the monthly usage credit applying to build usage, and at the contract rate on Enterprise. The changelog describes no change for existing teams.
Our take: this matters to anyone who opens a new Vercel team per client or per product. At $0.014 a minute, a 5-minute build costs 7 cents, so 1,000 such builds a month come to $70 at that rate (our arithmetic), previews included. Before you create the next team, count its builds per month, and turn on the preset below.
Pull-request-only preset for preview builds
On 17 November 2026 Vercel added a pull-request-only preset for preview builds (opens in a new tab): it skips preview builds for branches without an open pull request. In the project's Deployments > Configuration page, under Ignored Build Step, choose "Only build previews for pull requests" from the Behavior menu and save; production deployments and branches with an open pull request still build. On GitHub and GitLab, the skipped previews are cancelled before they use any Build CPU Minutes.
Our take: with builds billed by the minute, this is the cheapest setting to change this week. Two cautions: a preview link you send from a branch without a pull request stops appearing, so open a draft pull request when someone needs to see the work; and the preset fills in the ignore command itself, so copy any custom ignore script you have before you switch.
Vercel Blob deletes files by TTL
Vercel Blob supports TTLs (opens in a new tab) since 17 November 2026, with @vercel/ 2.9.0. Pass ttlDays, from 1 to 730, to put(), copy(), putImage(), multipart uploads or presigned upload URLs; the result and head() return expiresAt. setTtl() sets, changes or removes (ttlDays: null) the TTL of blobs you already have, for one URL or up to 1,000 at a time, without uploading them again; the CLI's vercel blob set-ttl needs Vercel CLI 65.16.0. Setting a TTL costs one Advanced Operation per started 60 days on top of the write, so a 7-day TTL costs one and the 730-day maximum 13; clearing one and the deletions are free.
Our take: a cron job that lists a folder and deletes old exports can go, and so can the bug where it fails silently and the bucket keeps growing. Set the TTL at write time for anything with a retention rule: exports, temporary uploads, files a user sent for one job. The TTL counts from the upload date, not from the last read, so it is the wrong tool for files that should live while people use them.
AI Gateway
Private Inference brings your own model endpoints
AI Gateway supports Private Inference (opens in a new tab) since 16 November 2026: you register an endpoint you run (a fine-tuned, self-hosted or prerelease model; Vercel's example is a vLLM server) as a private provider, and call its models as private/ through the same SDKs, APIs, keys and observability as catalog models. Only your team can see or call them, credentials are stored encrypted, Zero Data Retention can be set per provider, and registering from the command line needs Vercel CLI 65.16.0. The changelog doesn't state the price.
Our take: one gateway for the fine-tuned model and the catalog means one place for logs, spend and fallbacks. The optional fallback (providerOptions.gateway.models) keeps a feature up when your GPU box goes down, but it also sends that request to a catalog provider; if you self-host to keep data in-house, leave the fallback out or point it only at a provider whose retention you have checked.
Decision-1.1 replaces Decision-1
Microsoft's Decision-1.1 is on AI Gateway (opens in a new tab) since 17 November 2026 as microsoft/. It takes the same requests as Decision-1 and answers the same three kinds of question (a probability that something is true, a choice among categories, a score on your scale), and Vercel calls it more accurate, better calibrated and faster. In TypeScript it runs through experimental_decide in ai 7.0.128 or later.
Our take: the switch is one model ID, but "better calibrated" means the probabilities move, so a threshold you tuned on Decision-1 (flag above 0.8, escalate above 0.95) needs a rerun on your labelled cases before it goes live. Decision models took a quarter of Gateway requests in October, per our note on the AI Gateway Production Index, which is why a quiet threshold drift here is worth an afternoon.
GPT-6.2 Astra and Sol
OpenAI's GPT-6.2 Astra and Sol are on AI Gateway (opens in a new tab) since 17 November 2026, as openai/ and openai/. Astra is for complex multi-step work: changes in existing codebases, document templates and computer use. Sol is for everyday coding, writing and agent workflows. Both support fast mode and OpenAI's Ultrafast tier, and GPT-6.2 Astra's Ultrafast output price is 60% lower than GPT-6 Astra's.
Our take: if you pay for GPT-6 Astra on Ultrafast, the new version is the obvious test. The changelog gives no base prices, so check the model page before you move anything else.
Hy4 Flash at half price until 29 November
Tencent's Hy4 Flash (opens in a new tab) (tencent/) joined AI Gateway on 16 November 2026: the lower-cost model of the Hy4 family, with a 1M token context window, at roughly a tenth of Hy4's price per token. Input, cached input and output tokens are 50% off through 29 November 2026.
Our take: half price through 29 November is a good window to run your evals for bulk work (summaries, extraction, classification over long documents). Price the decision at the full rate, which is what you pay from 30 November.
MAI-Image-2.7 and MAI-Image-2.7-Flash
Microsoft's MAI-Image-2.7 and MAI-Image-2.7-Flash (opens in a new tab) are on AI Gateway in public preview since 17 November 2026, as microsoft/ and microsoft/. The first is for final assets, Flash for faster, cheaper drafts and variations. Both generate up to 2048 × 2048 pixels, take up to five reference images, and produce a transparent background with providerOptions.azure.background: 'transparent'; they need AI SDK 6 or later.
Our take: the transparent background is the useful part for a shop or a design tool, because it removes the background-removal step after generation. Explore prompts on Flash and render the final one on the full model.
Claude
Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 now carry the text watermark
The Claude release notes (opens in a new tab) of 16 November 2026 say that text from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 now carries Anthropic's text watermark on supported platforms. Per the text watermark page (opens in a new tab), it doesn't change the meaning, quality or readability of the output, adds no tokens or hidden characters and carries no information about you; if a request sets temperature, top_p or top_k to a non-default value, the output might not carry it. Anthropic's help article How Claude marks AI-generated content (opens in a new tab), which lists the models and platforms, covers the Claude API, the Claude apps and Claude Code, and the models on AWS, Google Cloud and Microsoft Foundry, with Amazon Bedrock taking a few extra days for some models; detection is in private preview for regulators, media, researchers and selected enterprises.
Our take: nothing in your code has to change, and there is nothing to strip in post-processing, because the mark isn't a character. If you need your output to carry it, omit the three sampling settings and check the request body your framework sends, since some set them for you. Neither page describes a way to turn it off.
What is eve 1.0 on Vercel?
eve is Vercel's open-source framework for building agents: it runs the loop that calls models, runs tools and delegates work to subagents, while you choose the model, the tools and which actions need approval. Version 1.0, announced on 17 November 2026 after a public preview that began in June, is the first generally available release; it needs Node.js 24 or newer, and the upgrade guide lists the breaking changes from the preview.
How much do Standard builds cost on Vercel?
For Pro and Enterprise teams created from 17 November 2026, a Standard build is billed by usage: $0.014 per build minute on Pro, with the monthly usage credit applying to build usage, and the contract rate on Enterprise. New teams no longer get an included Standard build slot. Vercel's changelog describes no change for existing teams.
How do I set a TTL on Vercel Blob?
Install @vercel/blob 2.9.0 and pass ttlDays, from 1 to 730, to put(), copy(), putImage(), a multipart upload or a presigned upload URL. For blobs you already have, setTtl() sets or changes the TTL for one URL or up to 1,000 at a time, and ttlDays: null removes it. The TTL counts from the upload date, and deletion is free.
How do I build Vercel previews only for pull requests?
In the project's Deployments > Configuration page, under Ignored Build Step, choose 'Only build previews for pull requests' from the Behavior menu and save. Production deployments and branches with an open pull request still build; on GitHub and GitLab, skipped previews are cancelled before they use Build CPU Minutes.
Does Claude watermark text?
Yes. Since 16 November 2026, text from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carries Anthropic's text watermark on supported platforms, as newer models already did. It adds no tokens or hidden characters and doesn't change the meaning or quality of the output. A request that sets temperature, top_p or top_k to a non-default value might not carry it; detection is in private preview.