Deferred to Q1 — 2026-10-26
Not part of the Q4 plan; stays in Backlog for Q1 planning.
Current review — 2026-10-20
The 27.3% baseline and top-20 table are a 2026-07-26 repository snapshot, not current production measurements. The catalogue is now database-backed (phaseoteam/Phaseo#2313); provider/fallback improvements do not prove the 40% exact-route target. Remaining: recompute active direct-provider redundancy from current canonical route/capability data, deduplicate identities, separate aggregator/family fallback and record live verification for activations. Keep the historical table as context only.
Objective
Raise exact model-capability route redundancy without hiding single-provider risk behind another gateway aggregator.
Snapshot generated from the repository catalogue on 2026-07-26:
- 51 providers with at least one active gateway route
- 472 active exact
model_id + capability_idroutes - 129 routes with 2+ active providers (27.3%)
- 343 single-provider routes (72.7%)
- 124 single-provider routes already have at least one catalogued but inactive alternative
An exact-route score is intentionally stricter than model-family fallback. Proprietary models may remain inherently single-provider unless Phaseo activates an authorized direct distributor.
Top-20 provider scorecard
| Provider | Active routes | Redundant | Exclusive | Redundancy |
|---|---|---|---|---|
| openai | 107 | 6 | 101 | 5.6% |
| novita | 91 | 72 | 19 | 79.1% |
| alibaba-cloud | 69 | 30 | 39 | 43.5% |
| deepinfra | 64 | 50 | 14 | 78.1% |
| siliconflow | 56 | 49 | 7 | 87.5% |
| atlascloud | 55 | 48 | 7 | 87.3% |
| gmicloud | 38 | 38 | 0 | 100% |
| venice | 35 | 32 | 3 | 91.4% |
| google-ai-studio | 29 | 9 | 20 | 31.0% |
| voyage | 28 | 0 | 28 | 0% |
| weights-and-biases | 26 | 24 | 2 | 92.3% |
| together | 23 | 19 | 4 | 82.6% |
| crofai | 22 | 17 | 5 | 77.3% |
| nebius-token-factory | 22 | 18 | 4 | 81.8% |
| mistral | 20 | 0 | 20 | 0% |
| fireworks | 17 | 14 | 3 | 82.4% |
| ionrouter | 17 | 15 | 2 | 88.2% |
| z-ai | 16 | 11 | 5 | 68.8% |
| venice-e2ee | 12 | 12 | 0 | 100% |
| anthropic | 11 | 3 | 8 | 27.3% |
Ruthless activation order
P0 — direct proprietary redundancy
- Activate and live-verify Amazon Bedrock for supported Claude routes currently exclusive to Anthropic.
- Activate and live-verify Google Vertex for supported Claude and Gemini routes.
- Activate current Amazon Bedrock alternatives for Mistral routes where the catalogue already records active capability support.
- Decide explicitly whether OpenAI exact-route redundancy may use Azure/direct cloud distribution. Do not count any third-party gateway aggregator as production redundancy without a written exception.
P1 — existing open-weight alternatives
- Add a second direct provider for Fireworks-exclusive Qwen embedding/rerank routes using the existing OVHcloud/Scaleway catalogue candidates.
- Add a second direct provider for DeepInfra-exclusive open-weight routes where NVIDIA, Cloudflare, GitHub Models, or Bedrock candidates already exist.
- Review Novita-exclusive routes and consolidate the duplicate
novita/novita-aiidentity before treating that pair as redundancy. - Promote already-configured deranked alternatives only after contract and live checks prove equivalent model identity.
P2 — portfolio policy
- Mark inherently proprietary single-provider routes as accepted risk, not “missing work.”
- Define model-family fallback presets for accepted-risk routes.
- Publish a scheduled scorecard that separates direct-provider redundancy, aggregator fallback, and family fallback.
Acceptance gates for every activation
- exact canonical model identity and capability match
- current pricing with source and effective window
- provider executor resolves for the capability
- non-stream, stream, tool, structured-output, and multimodal contracts as applicable
- live smoke passes with provider attribution and request-child observability
- routing is active, not merely catalogued
- rollback/derank path documented
- scorecard shows the route moving from 1 to 2+ direct active providers
Target
Move direct exact-route redundancy from 27.3% to at least 40% without activating unverified routes or masking risk through another aggregator.
Created with Codex
Activity
10
opencode-agent[bot]3mo agoIssue understanding
This is a [Chore] issue tracking a systematic effort to increase exact-route provider redundancy from 27.3% to 40% across the Phaseo gateway catalogue. The work involves activating verified alternative providers for routes currently served by a single provider, with strict acceptance gates for each activation. The issue provides a prioritized checklist (P0-P2) covering proprietary redundancy (Bedrock for Claude, Vertex for Gemini/Claude), open-weight alternatives, and portfolio policy decisions.
The catalogue contains 472 active exact routes with 343 single-provider routes (72.7%). The goal is to activate enough verified alternatives to reach 40% redundancy while maintaining exact model identity and capability matches.
AI Agent Prompt (Copy/Paste)
This issue tracks raising exact-route provider redundancy from 27.3% to 40% across the Phaseo gateway catalogue. The work involves activating verified alternative providers for single-provider routes following strict acceptance gates.
Key areas to investigate:
1. Catalogue data and route definitions in `packages/data/catalog/src/data` - understand current route structure and single-provider routes
2. Provider configurations and routing logic in `apps/api` - understand how routes are activated and verified
3. The acceptance gates criteria listed in the issue - ensure any changes meet model identity, pricing, and observability requirements
Start by:
1. Examining the catalogue data structure to understand how routes and providers are defined
2. Identifying which routes currently have inactive alternatives already catalogued (124 routes mentioned)
3. Reviewing P0 priorities: Bedrock for Claude routes, Vertex for Claude/Gemini, and Mistral alternatives
4. Checking existing provider activation patterns to understand the verification process
The issue provides a detailed top-20 provider scorecard and prioritized activation order. Use this to focus efforts on high-impact areas like OpenAI (5.6% redundancy), Anthropic (27.3%), Google AI Studio (31.0%), and Voyage (0%).Clarifying Questions (Optional)
- Should the agent focus on P0 proprietary redundancy first (Bedrock/Vertex), or tackle P1 open-weight alternatives where more catalogue candidates already exist?
- Is there an existing verification/smoke-test framework for live provider activation, or does this need to be established?
PHA-311