Alibaba Token Plan Wiki

Public, measured reference data for Alibaba Cloud and Qwen Cloud Token Plan.

This repository covers:

• the Global/Singapore and China/Beijing gateways;
• Personal and Team model entitlements;
• chat, image, video, and audio (TTS/realtime) catalogue data;
• OpenAI-compatible and Anthropic-compatible behaviour;
• model capability probes;
• Token Plan credit pricing.

The data is measured against the live gateways because vendor pages and community catalogues can lag or disagree with runtime behaviour.

Start here

Resource	Purpose
Machine-readable catalogue	Sanitised current model, tier, modality, limit, and capability data
Interactive pricing chart	Compare measured Token Plan credit costs
Credit-rate experiment	Method and raw reasoning behind 1 credit = 0.01 RMB
Mainland pricing reference	Chinese Mainland model pricing
International pricing reference	International pricing and capability comparisons
Raw probe results	Redacted measurements used by the analysis

Current catalogue

As of 18 August 2026:

• Team exposes 17 chat models; Personal exposes 8 (a strict subset).
• /models lists 22 ids per Team key: 16 chat + 4 image + 1 TTS + 1 realtime audio. The 3 video models never appear on /models (async endpoint).
• One additional chat model (deepseek-v4-pro-0813) is servable by exact id but not surfaced on /models; it is in the catalogue as unlisted.
• The exact counts move as models land. data/models.json carries the current list and its updated date; trust it over this prose.
• Both tiers use the same sk-sp- key prefix, so the tier cannot be inferred from the key.
• Global and China use separate credentials and endpoints.
• A successful /models response is catalogue evidence, not proof that a subscription can perform inference. A lapsed subscription may still list models and then deny calls.

The complete current list and capability matrix live in data/models.json. Downstream integrations should cite a tag or commit rather than silently copying an unversioned list.

Endpoints

Region	OpenAI-compatible base URL	Anthropic-compatible base URL
Global	https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1	https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic
China	https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1	https://token-plan.cn-beijing.maas.aliyuncs.com/apps/anthropic

Token Plan keys are separate from DashScope pay-as-you-go and Alibaba Coding Plan credentials.

Pricing result

1 Token Plan credit = 0.01 RMB.

The gateway burns credits using Chinese Mainland model prices. For example, a model priced at 1 RMB input and 2 RMB output per million tokens consumes about 100 input credits and 200 output credits per million tokens.

The measured experiment, alternative hypotheses, and raw numbers are in data/credit-rate-experiment.md.

Reproducing probes

The scripts read credentials only from the environment. They never print the key.

export ALIBABA_TOKEN_PLAN_API_KEY=sk-sp-...
python3 probe-credit-rate.py --dry-run
python3 probe-capabilities.py --help

Live probes may consume plan credits. Before running one:

1. Read the script and confirm the endpoint, models, call count, and token bounds.
2. Obtain the key owner's express permission.
3. Make sure the owner understands the likely credit cost.
4. Start with the smallest targeted probe that can answer the question.

Do not use an AI agent to run a catalogue-wide or token-heavy probe without that informed permission.

Evidence rules

• Runtime claims should identify the endpoint, date, model, parameters, and observed result.
• A model-generated explanation is not evidence by itself.
• A 200 response does not prove a modality worked. Some text-only models silently ignore unsupported media.
• A validation error can reveal accepted parameter values, but it does not necessarily reveal the model's real output or context ceiling.
• Missing regional or tier credentials should be reported as untested, not guessed.
• Never publish API keys, account identifiers, private paths, workspace IDs, or unredacted logs.

Downstream integrations

• Hermes Alibaba Token Plan plugin
• Hermes Agent
• models.dev

The public catalogue is generated from a more detailed working dataset and then privacy-scanned before publication. Corrections are welcome through a focused issue with reproducible evidence.

Scope of use

Token Plan is licensed for interactive use with compatible AI coding and agent tools. It is not intended as a general application backend, unattended batch processor, load-testing service, or shared key pool.

Licence

MIT. See LICENSE.