AI Gateway virtual models: OP as one Claude Code and Codex slug, Mid on Cursor

Verdict, October 30, 2026. OP as one vmc/ slug for a team that already sends Claude Code and Codex through AI Gateway and wants the model, the provider list, fallbacks, and retention in one place. Mid on Cursor. Its gateway page has you add creator/model-name IDs by hand, and its bring-your-own-key limits leave Tab and some Agent routes off the gateway. Mid when a request still needs its own provider order, tags, or a free-provider guard. Those lose once the slug sets the same field. No letter moves on the October 15 AI Coding Agents board. S stays locked. OP or Mid has not created a virtual model and has not called a vmc/ slug.

Pages below were read October 30, 2026. The model count in the launch post is Vercel's.

What opened on October 28

NEW On October 28, 2026, Vercel opened virtual models on AI Gateway. A virtual model is a saved configuration with an ID such as vmc/coding-agent. The caller sends that ID. AI Gateway applies the model, routing, fallbacks, region, compliance settings, provider options, and tags stored on it. An edit reaches new requests within a few minutes, with no redeploy and no per-developer config change. The slug cannot be renamed. Agent catalogs list language-model slugs only. Sources: changelog (opens in a new tab), launch post (opens in a new tab), docs (opens in a new tab).

VENDOR CLAIM The launch post says 125 new models from more than 10 labs became available on AI Gateway in the prior 90 days, about 10 a week. It names no list and no method. Source: launch post (opens in a new tab).

The changelog's create example is:

vercel ai-gateway virtual-models create coding-agent \
  --model anthropic/claude-opus-5.5 \
  --config '{"speed":"fast","zeroDataRetention":true}'

The slug Claude Code and Codex can select

BUFF Claude Code and Codex send a model ID and usually cannot attach provider order, fallbacks, region, or zero data retention. The slug carries those. vercel ai-gateway setup connects the agents. Claude Code lists it as claude-code/vmc/, plus [1m] when the underlying window is 1 million tokens or more. That page says the /claude-code prefix is a display ID, stripped before routing, and the bill follows the underlying model. Codex takes the slug in /model, as codex --model vmc/coding-agent, or as model = "vmc/coding-agent" in ~/.codex/config.toml. Restart an open session to see a new slug. A slug the agent already stored updates new requests within a few minutes. Vercel's edit example moves vmc/coding-agent to anthropic/claude-fable-5.1. Sources: virtual models (opens in a new tab), Claude Code (opens in a new tab), Codex (opens in a new tab).

NERF OpenAI excludes Codex usage through API keys from the API credits on eligible ChatGPT subscriptions. Vercel says a gateway key does not change that exclusion and does not raise signed-in Codex limits. Source: Codex (opens in a new tab).

UPDATE The launch post points the same slug at openai/gpt-6.1-sol in its switch example. Provider options belong to one provider, so an Anthropic effort value stops applying on an OpenAI model. The post says to review those options, and to aim a second slug such as vmc/coding-agent-candidate at the new model and run evals before production moves. That eval step is Vercel's advice. Source: launch post (opens in a new tab).

What the slug overrules

BUFF On a field the slug carries, the slug wins and the request is ignored: provider order, the provider restriction, model fallbacks, fast mode, sort, service tier, prompt caching, provider timeouts, inference region on system credentials, required capabilities, and tags. Provider options merge key by key, and the slug wins on each key it sets. Zero data retention and blocking prompt training only tighten. If either side turns one on, it stays on. A team model allowlist checks the model behind vmc/, so a forbidden model stays blocked. A team provider allowlist still applies on top of the slug. Sources: docs (opens in a new tab), launch post (opens in a new tab).

NERF A field the slug leaves unset still comes from the request. @default, valid on sort, service tier, prompt caching, and provider timeouts, throws away the caller's value and uses the system default. Compliance fields reject it. Required capabilities replace the request's list, so a caller that sent has with free loses that guard when the slug sets its own list. Tags on the slug replace tags on the request. The launch post calls that replacement the way spend lands on the task, and it says the Virtual Models page shows requests, spend, and success rate per slug. Source: docs (opens in a new tab).

NERF Provider Order is a preference. Providers left off the list can still serve afterward. Provider Only is the lock: no fallback outside it, and the request fails when every listed provider fails. Inference region does not apply to your own key. Fallback from fast mode is on by default. Owners and Members can create a slug, any member can call it, and only the creator or a team Owner can edit, archive, remove, or restore it. Archive returns 404. CLI remove asks for no confirmation. Source: docs (opens in a new tab).

The bill

PRICE AI Gateway charges no markup and no platform fee on tokens. You pay the provider list price. Payment processing fees can still apply. Enterprise invoice billing has none. The free tier is a subset of models, a limit returns 429, and other models need purchased credits, which ends the monthly free credit. Bring your own key has no gateway markup, requires that paid tier, and a failed key retries on system credentials billed to the credits balance. Subscription API credits stay with the provider. Gateway budgets skip that spend. Source: pricing (opens in a new tab).

PRICE Per-request zero data retention is no additional cost, on Pro and Enterprise. Team-wide zero data retention is $0.10 per 1,000 requests, on the same plans. A slug can turn the setting on for its own traffic and cannot turn a team-wide setting off. The docs do not say which meter a slug flag uses. A team provider allowlist, which still sits on top of a slug, is $0.10 per 1,000 successful requests on Pro and Enterprise. The pricing page does not say a slug's Provider Only list or its tags use the free per-request meters. Source: pricing (opens in a new tab).

Cursor is the catch

NERF Cursor's endpoint is https://ai-gateway.vercel.sh/cursor/v1. After npx vercel@latest ai-gateway setup --agent cursor, you paste the key under Settings, Models, turn on Override OpenAI Base URL, and add model IDs yourself. The examples are anthropic/claude-opus-5 and openai/gpt-6-astra. The page does not mention vmc/. The virtual-model docs say another agent can pass vmc/ anywhere it accepts a gateway model ID. The bare /v1 URL returns a 400 on tools.0.type. Vercel attributes the rest to Cursor while the override is on: Cursor's own non-OpenAI models stop, Tab never uses a custom key, some Agent and Auto modes stay on Cursor's models, and keyed traffic still passes through Cursor's backend. A slug cannot govern a request that never arrives. Sources: Cursor (opens in a new tab), virtual models (opens in a new tab).

The board stays on October 15

A virtual model is a routing record. The October 15 AI Coding Agents list keeps Codex, Claude Code, and Cursor in A. S stays locked. The Frontier Models list does not gain a vmc/ row.

What to do today

On Claude Code and Codex, create one language-model slug, select it in both, and restart any open session so the catalog reloads.
Use Provider Only when the provider list must be closed. Use Provider Order when unlisted providers should still serve.
On Cursor, read the bring-your-own-key table before the base URL override. Add a model ID only if Tab and Cursor's own non-OpenAI models can stay off.
On Pro, price tokens at the provider list and team-wide zero data retention at $0.10 per 1,000 requests. Both zero data retention rows name Pro and Enterprise. The slug's own flag is unpublished between those meters.

The weekly patch is on beehiiv: subscribe free (opens in a new tab).

The OP or Mid Team

Sources

Checked October 30, 2026.