docs: add Model fallback configuration guide

[stephenvapiai]

What changed

Adds a new help center page, Model fallback configuration (fern/customization/model-fallback-plan.mdx), and wires it into fern/docs.yml under Documentation → Assistants → Model configurations, right after the existing Voice/Transcriber fallback pages.

Why (customer evidence)

Over the last 7 days, "Language Model Providers" was the top-volume Ask AI (docs.vapi.ai) topic (197 conversations this week vs. 127 the prior week vs. a ~131/week trailing-4-week baseline — an accelerating topic, not a one-off spike).

Most of that volume is customers successfully getting answers about which models each provider supports (OpenAI, Anthropic, Gemini, Groq, MiniMax, xAI, Mistral, etc.) — that content already exists and is accurate on the individual /providers/model/* pages, so it isn't the gap.

Within that same topic, a recurring cluster of distinct customers asked specifically about model-level fallback (fallbackModels / fallbackPlan on the model object) and the Ask AI assistant could not answer:

• "what triggers a switch to the next model in a fallbackModels list on the model object" — outcome: refused
• "about fallbackModels" — bot found no documentation for fallbackModels or fallbackPlan — outcome: refused

I verified against the live Create Assistant API reference that fallbackModels is a real, documented field — but only on OpenAIModel (list of fallback model IDs) and AnthropicBedrockModel (at most one same-provider fallback). No other provider schema exposes it. Meanwhile Vapi already ships dedicated guides for the analogous features on the other two pipeline components (voice-fallback-plan.mdx, customization/transcriber-fallback-plan.mdx) — there was no equivalent page for models, which is exactly the shape of confusion customers hit (they expected a fallbackPlan-style page like voice/transcriber have, and the assistant had nothing to cite).

This is a missing content gap, not a wording tweak: fallbackModels previously only appeared in changelog entries and a Simulations example, never in a standalone how-to guide.

Scope / what this intentionally does NOT claim

• Does not claim cross-provider model fallback exists (unlike voice/transcriber) — the new page explicitly calls this out so it doesn't create a different kind of confusion.
• Does not add Dashboard UI steps, since I could not confirm a Dashboard control for model-level fallback (voice/transcriber fallback pages document Dashboard steps I could verify; this one is API-only until confirmed otherwise).
• Scoped to the two providers (OpenAIModel, AnthropicBedrockModel) that actually expose fallbackModels in the live API schema — verified directly against the Create Assistant API reference, not assumed.

Excluded candidates (per review scope)

Other candidate gaps touching security, auth/authorization, data privacy, or compliance topics (e.g. a customer support request for architecture/security/data-retention documentation, and the separate Compliance Plans taxonomy bucket) were intentionally skipped for this PR and not drafted, per the review's exclusion criteria for that category.

Checks done

• Confirmed no recent commit or open PR already covers model fallback (list_commits on the two sibling fallback pages, search_code for fallbackModels, list_pull_requests).
• Pinned the edit to the current main tip (d9efddec9207511d8e6c249fbdad82db30f89f35) and rechecked main hadn't moved immediately before branching.
• Verified the fallbackModels schema (allowed values, per-provider availability) directly against /api-reference/assistants/create.

[stephenvapiai]

Why this gap, and what was excluded

• Source: Enterpret Context Graph, last 7 days (2026-10-05 to 2026-10-12) vs. prior 7 days vs. trailing 4-week baseline, across Zendesk-style support tickets, Discord, Ask AI (docs.vapi.ai) conversations, and NPS surveys.
• "Language Model Providers" was the top-volume, accelerating topic this week (197 vs. 127 prior week vs. ~131/week baseline). Within it, a recurring cluster of distinct customers asked about model-level fallbackModels/fallbackPlan and the Ask AI assistant returned refused/silent — i.e. no documentation to cite. Verified against the live Create Assistant API reference that fallbackModels is real but undocumented outside changelog entries, and only on OpenAIModel + AnthropicBedrockModel.
• Verified via list_commits (on voice-fallback-plan.mdx and transcriber-fallback-plan.mdx), search_code for fallbackModels, and list_pull_requests that no recent commit or open PR already covers this, to avoid duplicating in-flight work.
• Excluded from this PR by review scope: other candidate gaps this week that touched security vulnerabilities, authentication/authorization risk, privacy/data exposure, regulatory obligations, or compliance claims — e.g. a customer support request for architecture/security/data-flow/data-retention documentation, and the separate Compliance Plans taxonomy bucket — were intentionally not drafted or opened as PRs under this review's exclusion criteria.
• The diff is scoped to only the drafted content change: the new fern/customization/model-fallback-plan.mdx page and its single fern/docs.yml nav entry. No unrelated files were touched.