Research investigation R1118 / comparison
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.
Evidence cutoff: November 18, 2026. The collection records documentation snapshots retrieved at 2026-11-18T16:32:59.622Z, after the packet timestamp of 2026-11-18T16:30:57.426Z. The supplied timing note is truncated, so exact historical page versions at the packet timestamp are not established.
Version ledger
Frozen public editions
Each edition preserves the records, method, sources, and downloads available at publication time.
Evidence cutoff: November 18, 2026. The collection records documentation snapshots retrieved at 2026-11-18T16:32:59.622Z, after the packet timestamp of 2026-11-18T16:30:57.426Z. The supplied timing note is truncated, so exact historical page versions at the packet timestamp are not established.
- Dataset ID
- spd:perplexity-openrouter-or-vercel-which-failover-settings-preserve-the-requested-model-a93e8355
- Stable URL
- /research/perplexity-openrouter-or-vercel-which-failover-settings-preserve-the-requested-model-a93e8355
- Version
- v1
- Coverage
- Live collection
- Records
- 2
- Fields
- 7
- Updated
Measurement technique
How to read this report
- 01Reused the supplied evidence, including saved documentation snapshots and the collection's reported direct page checks. Recovering this material is not a new collection or experiment.
- 02Bounded the sample to seven official records covering three gateways. Compared a fixed, explicitly named catalog model without aliases, virtual models, automatic model routers, application retries or SDK retry policies. Evaluated optional alternate-model configurations separately.
- 03Planned an evidence_matrix with gateway/configuration rows and columns for upstream rate limits, provider timeouts, eligible-provider exhaustion, provider restrictions, model substitution, terminal behavior and source. Classify outcomes as model-preserving, model-changing, terminal error or unspecified; distinguish defaults from explicit settings.
- 04Perplexity matrix entry: retry another deployment of the same model on provider errors, upstream rate limits or timeouts when another deployment exists; return HTTP 429 with Retry-After after deployment exhaustion. No per-provider configuration is documented. Invalid requests and context overflow return HTTP 400 without another deployment attempt; streams failing after output begins end with an in-band error. Source: https://docs.perplexity.ai/docs/orchestrator/routing-and-reliability
- 05Perplexity model-selection entry: perplexity/auto is optional preview model selection and can change models between turns. The sampled records do not establish a failure-triggered alternate-model chain. The separate technical account reports internal measurements for October 4–November 2, 2026, not a matched cross-gateway experiment. Source: https://www.perplexity.ai/cs/hub/blog/orchestrator-one-api-for-reliable-inference
- 06OpenRouter matrix entry: provider.allow_fallbacks defaults true; order prioritizes providers and only restricts eligibility. With allow_fallbacks=false and no order, only the top-ranked endpoint remains. With nonempty order, unlisted public providers are excluded but eligible private endpoints remain. An explicit models list permits substitution on errors; unsuccessful fallback returns an error. Default model partitioning prioritizes primary-model endpoints, but provider.sort.partition='none' on
Sources
Evidence
3 publishers supporting 2 records. Expand a publisher to inspect its cited pages.