Three AI gateways preserve model choice during retries—but their fallback rules differ
Original analysis: Perplexity, OpenRouter and Vercel document same-model recovery. Provider restrictions and alternate-model settings determine how far that recovery can go.
All three sampled AI gateways document same-model failover, but fallback controls differ
Same-model failover is documented across all three gateways. The useful comparison is which providers remain eligible, when another model becomes permissible and what happens when eligible providers fail—not a reliability ranking.
Explore the full research
Superpower DailyOriginal research
Three AI gateways preserve model choice during retries—but their fallback rules differ
Listen to this story
The audio brief
About 1:33
0:001:33
Story brief
3 key points
For teams evaluating gateway controls, the key distinction is what happens after a route fails: provider ordering may not limit eligibility, and a configured backup model can change the response model and its price. An official-document review, with a cutoff of November 18, 2026, finds that Perplexity Orchestrator, OpenRouter and Vercel AI Gateway all document retries that can keep an explicitly requested catalog model. The comparison does not measure live reliability, latency, cost or successful recovery.
01
OpenRouter’s `order` field prioritizes providers but does not exclude other eligible providers; `only` sets a strict eligibility list.
02
With OpenRouter `allow_fallbacks=false` and a nonempty order list, unlisted public providers are excluded, but eligible private endpoints may remain.
03
OpenRouter model fallbacks can follow downtime, rate limits, moderation flags or context-length validation errors; pricing reflects the model ultimately used.
Switching AI providers after a failed request does not have to mean switching models. Superpower Daily’s public-document review finds that Perplexity Orchestrator, OpenRouter and Vercel AI Gateway all describe retries that preserve the requested model. Their meaningful differences lie elsewhere: which providers remain eligible, when another model may take over, and when recovery stops.
That finding narrows what same-model failover can establish as a competitive distinction. All three sampled services document it; none should be treated as uniquely capable on that basis. But the shared feature does not make their controls interchangeable. A preferred-provider list can leave other destinations available, while an explicitly configured backup-model list can authorize a different model to answer.
Original analysis: the comparison’s boundaries
We compared official documentation for the three gateways, with an editorial cutoff of November 18, 2026. The baseline was an explicitly named catalog model—not an alias, virtual model or automatic model selector. Optional alternate-model settings were examined separately. Application retries and software-library retry policies were outside the comparison. This is a review of documented behavior, not a live reliability experiment, and it does not rank availability, latency, cost or successful recovery.
The model can stay while the host changes
The first distinction is between the model being requested and the deployment serving it. Perplexity’s Routing & Reliability documentation separates these decisions. Naming a catalog model calls it directly; provider routing then selects a deployment using observed error rates, capacity and latency. Provider errors, upstream rate limits or timeouts can trigger another attempt through a different deployment of that same model, when one exists.
OpenRouter also separates backup providers from backup models. Its provider.allow_fallbacks setting defaults to true, permitting backup providers when the primary is unavailable. Vercel documents the same basic model-preserving path: a model available through multiple providers can move to another provider if an attempt fails. Vercel’s default provider selection uses recent uptime and latency.
A preference list is not a boundary
OpenRouter’s Provider Routing documentation makes a consequential distinction between order and only. The order field prioritizes providers; it does not, by itself, exclude other eligible providers. The only field supplies a strict eligibility list. A configuration that says which hosts to try first therefore answers a different question from one that says which hosts may receive the request at all.
Disabling OpenRouter fallbacks adds an exception worth reading closely. With allow_fallbacks=false and no order list, only the top-ranked endpoint remains. With a nonempty order list, unlisted public providers are excluded, but eligible private endpoints remain. The setting therefore does not universally promise one provider attempt. Its effect depends on the accompanying configuration and the endpoints available.
Vercel’s filtering and ordering documentation draws a similar boundary. Providers omitted from order remain available after the named providers; only restricts eligibility. Without model fallbacks, a request fails when the allowed providers cannot serve it, including because they are down, rate-limited or out of capacity. Perplexity takes a different approach: its sampled documentation offers no per-provider configuration.
Changing models is a separate permission
OpenRouter’s Model Fallbacks documentation describes an explicit models list that allows substitution. Errors that can trigger it include downtime, rate limits, moderation flags and context-length validation failures. If the fallback also fails, OpenRouter returns an error. Requests are priced using the model ultimately used, which appears in the response’s model field. A backup model is thus also a possible change in the bill.
There is an important qualification to the apparent sequence. OpenRouter normally groups endpoints by model, prioritizing the primary model’s endpoints. Setting provider.sort.partition to none removes that grouping and permits sorting across endpoints for all listed models. Under that optional configuration, the primary model’s position is no longer the same assurance that its endpoints will be tried first.
Vercel’s alternate-model configuration is providerOptions.gateway.models. Its documented sequence applies provider routing rules to each model, then advances to the next model after all eligible providers fail. The first successful model/provider combination supplies the response. This explicitly connects the two controls: a provider restriction can exhaust the available routes for one model before the gateway moves to its configured backup.
Perplexity’s optional perplexity/auto selector should not be confused with that recovery chain. It selects a model for each request, and different conversation turns can use different models. The sampled records do not establish a failure-triggered alternate-model chain after deployment exhaustion. That is a limit on this comparison’s evidence, not proof that an undocumented capability cannot exist.
Recovery also has a clock—and a stopping point
Vercel documents explicit per-provider timeout controls under providerTimeouts.byok, but they apply only to bring-your-own-key credentials. When the configured timeout expires, the gateway aborts the attempt and tries the next available provider. For streaming responses, the timer clears at the first chunk; for non-streaming requests, it runs until the complete response arrives. These are different waiting limits, not a universal promise of recovery throughout a response.
The documented distinction is not unique same-model failover: all three support it. The material differences are provider restrictions, timeout controls, explicit model-substitution settings and terminal-error details.
Vercel: default routing chooses providers using recent uptime and latency. providerOptions.gateway.order controls same-model provider preference but leaves unlisted providers eligible; only restricts eligibility. Without model fallbacks, exhausted allowed providers cause failure, including when down, rate-limited or out of capacity.
Perplexity: naming a catalog model enables automatic same-model deployment retries on provider errors, upstream rate limits or timeouts when another deployment exists. Every deployment failing returns HTTP 429 with Retry-After, not a different model.
Reader comments
Newest comments first. Replies stay oldest first.