Stale Alibaba Token Plan catalogs: phantom glm-5.3 rows, wrong qwen3.8-max window, missing callable ids

[oliver-mee]

Client or integration

Other (provider catalog data; surfaces in the Codex model picker)

Summary

Both alibaba-token-plan presets (Beijing and International) ship catalog facts that do not match what the Token Plan gateway actually serves. Four classes of defect, all re-verified on 2026-09-07 and 2026-09-09 against the live gateway on both regions and both subscription tiers:

1. Phantom rows. glm-5.3 and glm-5.3-flash are listed on both Token Plan presets. These models exist on Z.AI endpoints, but the Token Plan gateway has never served either id; selecting the row fails with model-not-found. They arrived via the 260826 seed commit that propagated GLM-5.3 across every GLM-carrying catalog, including the plan presets.
2. Wrong context window. qwen3.8-max is pinned to 983_616 in both presets. That number is the CLAUDE_CODE_MAX_CONTEXT_TOKENS client setting, not a model window; the real window is 1,000,000. The same figure circulated through two other third-party catalogs before, so it is easy to re-import by accident.
3. Missing callable ids. qwen3.8-flash (GA, listed by /models on both tiers), deepseek-v4-pro-0813 and deepseek-v4-flash-0731 (callable snapshots; the 0813 one is notably absent from the gateway's own /models listing) are not in the presets. And deepseek-v4-pro was recently dropped from the international lineup, but the plan gateway still lists and serves it on both regions (this is the plan roster, not the DeepSeek API, so upstream retirement does not apply).
4. Stale windows and missing ceilings. glm-5/glm-5.1 are 202,752 (not 1M) and MiniMax-M2.5 is 196,608 (not 204,800). Both presets also ship no modelMaxOutputTokens at all, so nothing clamps or advertises the output ceilings (deepseek-v4 family 393,216 on this gateway, probed at the boundary).

Expected: selecting any row in the Token Plan pickers routes to a model the subscription can actually call, with windows and output ceilings that match the gateway.

Reproduction

1. Configure the alibaba-token-plan-intl preset with any Token Plan key (Personal or Team tier).
2. In a routed client, select glm-5.3 from the model list.
3. The request fails at the gateway: the id is not in the plan's served set. On the free /models surface (probed 2026-09-07 and 2026-09-09, both regions, both tiers) Token Plan lists only glm-5, glm-5.1 and glm-5.2, and the vendor's own plan overview page lists the same three.
4. Same for a qwen3.8-flash request that is never offered at all: it is callable and listed by the gateway /models on both tiers, but it is not in the preset, so liveModels: false static users cannot select it.

Version

2.51.0 (also present in 2.31.0 through 2.46.0)

Operating system

Ubuntu 24.04 (arm64)

Provider and model

alibaba-token-plan-intl / glm-5.3, qwen3.8-max, qwen3.8-flash, deepseek-v4-pro-0813

Logs or error output

# GET /compatible-mode/v1/models on a Team key and a Personal key, SG and Beijing (2026-09-09):
glm ids returned: glm-5, glm-5.1, glm-5.2   (no glm-5.3, no glm-5.3-flash)
# qwen3.8-flash IS returned on both tiers but is absent from the preset.

Screenshots and supporting files

No UI involved. Per-value evidence (gateway probes with dates and methods, including max_tokens boundary probes for every output ceiling) is maintained in the machine-readable catalog at https://github.com/oliver-mee/alibaba-token-plan-wiki

Redacted configuration

{
  "providers": {
    "alibaba-token-plan-intl": { "adapter": "openai-chat", "authMode": "key", "liveModels": false }
  }
}

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.

[lidge-jun]

리뷰 · 우선순위 58 / 80

이 이슈 #4787은 Alibaba Token Plan 카탈로그가 라이브 게이트웨이와 어긋난 네 가지 결함 클래스를 정리한다. 지금 dev HEAD(5e3029e6f, package 2.57.0)를 보면 src/providers/registry/model-seeds.ts의 ALIBABA_TOKEN_PLAN_MODELS/ALIBABA_INTL_TOKEN_PLAN_MODELS에 여전히 glm-5.3/glm-5.3-flash가 들어가 있고, src/providers/registry/entries-extended.ts의 두 프리셋은 qwen3.8-max를 983_616으로 두며, glm-5/glm-5.1을 1M으로, MiniMax-M2.5를 204_800으로 적고, modelMaxOutputTokens와 promptCacheKey는 Token Plan 쪽에 없다. 픽커에서 유령 행을 고르면 model-not-found가 나고, 창/출력 상한이 틀리면 클라이언트 쪽 1M·토큰 예산 계산도 틀린다. 재현 단계(프리셋 설정 → glm-5.3 선택 → 실패, qwen3.8-flash는 목록에조차 없음)와 2026-09-07/09 양 지역·양 티어 프로브 서술이 구체적이라 추적 이슈로 가치가 있다.

이미 수정 PR #4788(draft)이 Closes #4787로 열려 있다. 그 PR은 공유 창/출력/비전/preserve 상수로 두 지역을 맞추고, 유령 행 제거·1M 교정·누락 id·출력 상한·promptCacheKey까지 한 번에 넣는다. Z.AI 카탈로그의 glm-5.3는 의도적으로 남긴다. 다만 #4788은 아직 draft이고, tests/providers/provider-registry-parity.test.ts의 베이징 고정값이 구카탈로그를 잠그고 있어 그대로면 CI가 깨질 가능성이 크다. 이슈 본문이 미룬 per-tier(Personal vs Team) 분리와 liveModels 유지 판단도 메인테이너가 같이 봐야 한다.

라인 - src/providers/registry/model-seeds.ts ALIBABA_TOKEN_PLAN_MODELS / ALIBABA_INTL_TOKEN_PLAN_MODELS (현재 dev) - glm-5.3/glm-5.3-flash 유령 행이 이슈 클래스 1과 일치한다
라인 - src/providers/registry/entries-extended.ts alibaba-token-plan / alibaba-token-plan-intl modelContextWindows - qwen3.8-max: 983_616, glm-5/glm-5.1 1M, MiniMax-M2.5: 204_800이 이슈 클래스 2·4와 일치한다
라인 - 같은 엔트리에 modelMaxOutputTokens 부재 - 출력 상한 미광고/미클램프(이슈 클래스 4)
라인 - 누락 id qwen3.8-flash, deepseek-v4-pro-0813, deepseek-v4-flash-0731 및 intl에서 빠진 deepseek-v4-pro - 이슈 클래스 3
라인 - 수정 레일 #4788 - 방향은 맞지만 parity 테스트 미갱신·draft readiness 미완이 남아 이슈를 아직 닫을 수 없다

메인테이너의 판단이 필요한 지점

• #4788을 이 이슈의 단일 수정 레일로 확정할지(추천: 예). 다른 중복 패치는 현재 보이지 않는다
• Personal/Team 티어별 명단 분리를 이 이슈 범위에 넣을지, 후속 이슈로 쪼갤지(작성자는 후속으로 명시)
• 외부 프로브 위키만으로 증거 충분한지, in-repo 요약 커밋을 요구할지

너의 추천
KEEP. 이슈 서술은 dev 실제 카탈로그와 잘 맞고 재현도 분명하다. 수정은 #4788 하나로 모으고, #4788이 parity·providers CI를 맞춘 뒤 ready·머지되면 Closes #4787로 이슈를 닫으면 된다. 지금은 이슈를 닫지 말고, per-tier 후속만 별도 이슈/라벨로 남겨 두라. types/config 분할 close-don't-rebase 해당 없음.

이 댓글은 grok-bot이 작성했습니다

[lidge-jun]

Tracking: draft PR #4788 refreshes these catalogs against the live gateway.

Still present on dev (2b19983bfd): the phantom GLM rows at src/providers/registry/model-seeds.ts:440-466 and qwen3.8-max at 983_616 in src/providers/registry/entries-extended.ts:727-765.